1

Artificial Intelligence for Atmospheric Sciences:

A Research Roadmap
Martha Arbayani Zaidan, Senior Member, IEEE, Naser Hossein Motlagh, Senior Member, IEEE,
Petteri Nurmi, Tareq Hussein, Markku Kulmala, Tuukka Petäjä, Sasu Tarkoma, Senior Member, IEEE

Abstract—Atmospheric sciences are crucial for understanding Similarly, devastating floods in Spain in 2024 were intensified
environmental phenomena ranging from air quality to extreme by excessive heat in the Mediterranean and Atlantic Oceans,
weather events, and climate change. Recent breakthroughs in exemplifying the interconnectedness of rising sea temperatures
sensing, communication, computing, and Artificial Intelligence
arXiv:2506.16281v1 [cs.ET] 19 Jun 2025

(AI) have significantly advanced atmospheric sciences, enabling and extreme weather [3]. These and numerous other related
the generation of vast amounts of data through long-term Earth events highlight the accelerating effects of climate change
observations and providing powerful tools for analyzing atmo- and reinforce the need for immediate action [4]. Accurate
spheric phenomena and predicting natural disasters. This paper monitoring of atmospheric conditions, from local weather
contributes a critical interdisciplinary overview that bridges the patterns to global climate models, is vital for addressing
fields of atmospheric science and computer science, highlighting
the transformative potential of AI in atmospheric research. these challenges as it enables policymakers, scientists, and
We identify key challenges associated with integrating AI into industry leaders to make informed decisions that impact both
atmospheric research, including issues related to big data and the environment and society [5].
infrastructure, and provide a detailed research roadmap that Recent breakthroughs in sensing, communication, comput-
addresses both current and emerging challenges. ing, and Artificial Intelligence (AI) have significantly advanced
atmospheric sciences by facilitating the generation of extensive
I. I NTRODUCTION datasets through long-term Earth observations and providing
powerful tools to better understand atmospheric phenomena

U Nderstanding the atmosphere and environment is crucial

for sustaining life on Earth. Atmospheric and envi-
ronmental sciences play a key role in mitigating climate
and predict events such as natural disasters. However, these
advancements also present new challenges. Firstly, the rapid
growth of big data from ground measurements, satellite ob-
change, predict natural disasters, and ensuring the health of servations, laboratory experiments, and Internet of Things
humans and ecosystems [1]. This importance is highlighted (IoT) sensor networks is overwhelming conventional analysis
by the increasing frequency of severe weather events and tools [6], [7]. Satellite remote sensing alone produces vast
environmental degradation, which together pose a looming amounts of high-dimensional data, while numerical simula-
threat to global stability [2]. For instance, November 2024 tions of weather patterns and environmental changes introduce
was recorded as the second-warmest November globally, with computationally intensive datasets [8]. Similarly, the prolifer-
temperatures averaging 1.62°C above pre-industrial levels. ation of low-cost IoT sensor networks has dramatically in-
creased the availability of real-time atmospheric data, compli-
This project is supported by the Research Council of Finland through the
University Profiling funding initiative InterEarth (Grant No. 353218) and an cating data management and analysis [9], [10], [11]. Secondly,
Academy Research Fellowship (Grant No. 355330). In addition, financial maintaining the extensive infrastructure required for various
supports through the Research Council of Finland via the Atmosphere, Climate sensing, communication, and computing platforms presents
Competence Center (project numbers 337549, 357902, 359340) and via
Research Infrastructure of Institute for Atmospheric & Earth System Research its own challenges. Ensuring the accuracy, reliability, and
- INAR RI (367739) are gratefully acknowledged. availability of these systems requires continuous calibration,
M.A. Zaidan is with Department of Computer Science and Institute for data synchronization, and fault management [12], particularly
Atmospheric Science and Earth System Research (INAR), University of
Helsinki, Finland (e-mail: [email protected]). in remote or extreme environments, where wear and calibration
N. H. Motlagh is with Department of Computer Science, University of drift can degrade data quality [13].
Helsinki, Finland (e-mail: [email protected]). Addressing these challenges requires advanced computing
P. Nurmi is with Department of Computer Science and Helsinki Institute
of Sustainability Science (HELSUS), University of Helsinki, Finland (e-mail: resources and innovative Artificial Intelligence (AI) method-
[email protected]) ologies. High Performance Computing (HPC) is essential for
T. Hussein is with Institute for Atmospheric Science and Earth processing, storing, and analyzing vast datasets in real-time,
System Research (INAR), University of Helsinki, Finland (e-mail:
[email protected]). whether from high-dimensional satellite imagery, complex
M. Kulmala is with Institute for Atmospheric Science and Earth simulations, or large-scale sensor networks [14]. AI-driven ap-
System Research (INAR), University of Helsinki, Finland (e-mail: proaches are revolutionizing atmospheric infrastructure man-
[email protected]).
T. Petäjä is with Institute for Atmospheric Science and Earth agement by automating critical tasks such as sensor cali-
System Research (INAR), University of Helsinki, Finland (e-mail: bration, anomaly detection, and predictive maintenance [15].
[email protected]). These innovations enhance the efficiency and reliability of dis-
S. Takoma is with Department of Computer Science, University of Helsinki,
Finland (e-mail: [email protected]). tributed monitoring systems while reducing operational costs
(Corresponding authors: M.A. Zaidan.) and human intervention [16]. In parallel, AI, defined broadly as
 2

the simulation of human intelligence processes by machines, (B) Computational and Analytical Platforms, such as high-
is revolutionizing many scientific fields [17], including at- performance computing (HPC) systems, cloud-based plat-
mospheric sciences. AI encompasses a range of techniques, forms, and advanced data analytics tools [21]. These
from machine learning algorithms to deep learning methods platforms enable the processing of large-scale datasets,
capable of identifying complex relationships in large datasets, complex simulations, and the integration of AI-driven
making these AI-driven methods particularly well-suited for methodologies into atmoshperic sciences.
atmospheric research. These methods are particularly well- Access to sufficiently sophisticated sensing and computing
suited for processing the highly varied and voluminous data infrastructures is indispensable for advancing of atmospheric
that characterize atmospheric and environmental research. For research, particularly in the AI era, where studies require
example, by applying AI, researchers can optimize weather substantial computational power and extensive datasets [22].
prediction models, improve climate simulations, and derive These infrastructures provide decentralized resources for both
actionable insights from satellite imagery or sensor data [18]. data collection and computation. The widespread availability
This paper contributes a critical interdisciplinary overview of sensing systems enables fine-grained, high-resolution mea-
that bridges the fields of atmospheric science and computer surements, ensuring a comprehensive coverage of atmospheric
science, highlighting the transformative potential of AI in phenomena. Meanwhile, computational resources support low-
atmospheric research. We identify and analyze key challenges latency analysis of large atmospheric datasets. We next discuss
posed by big data and extensive infrastructure, providing a these two categories of infrastructure and their transformative
comprehensive assessment of current AI-driven methodolo- potential for atmospheric sciences.
gies and solutions that enhance data analysis and predic-
tive modeling. In addition, we outline a research roadmap
that emphasizes emerging trends such as edge computing,
AI-accelerated modeling, and interdisciplinary collaboration,
fostering synergies between atmospheric science and com-
putational technologies. By clarifying these intersections and
offering practical pathways forward, this work aims to catalyze
further advancements in environmental monitoring, disaster
prediction, and climate change mitigation efforts, contributing
to more sustainable interactions with our planet’s atmosphere.
The paper is organized as follows: in Section II, we re-
view the state-of-the-art atmospheric research infrastructure,
identify existing challenges, and explore AI-driven solutions
that enhance data analysis and predictive modeling. Section III
explores various branches of atmospheric sciences, analyzing
current practices, challenges, and potential advancements in
AI for areas such as air quality monitoring, operational mete-
orology, satellite remote sensing, and Earth system modeling.
This section highlights key applications of AI while providing
context on related research fields in environmental monitoring.
In Section IV, we present a research roadmap that outlines
future directions for AI in atmospheric sciences, emphasizing
emerging trends and opportunities for interdisciplinary col-
laboration. Finally, Section V summarizes the key insights
from the paper, reinforcing the transformative role of AI in
advancing atmospheric sciences.

II. I NFRASTRUCTURE FOR ATMOSPHERIC M ONITORING

AND A NALYTICS
Fig. 1: Atmospheric science infrastructure consists of (A)
Reliable infrastructure is crucial for the advancement of
Observation and Sensing Systems and (B) Computational and
atmospheric sciences, as it facilitates the seamless acquisi-
Analytical Platforms.
tion, analysis, and dissemination of essential environmental
data [19]. As illustrated in Figure 1, this infrastructure can be
broadly categorized into two key components:
(A) Observation and Sensing Systems, which include A. Observation and Sensing Systems
ground-based research stations, in-situ monitoring instru- As illustrated in Figure 1 (the lower part), observation and
ments, and satellite networks [20]. These systems are sensing systems can be categorized into ground-based and
designed to collect high-quality, continuous, and spatially remote sensing approaches.
diverse data essential for atmospheric studies and to Ground-based observations are primarily located near
provide the inputs needed for AI techniques. the Earth’s surface and include both mobile and continuous
 3

observation platforms. Mobile observation platforms include

vehicles, pedestrians, or other mobile units that transport data
collection units, as well as fixed sensors that may be relocated
periodically. These extend measurements from a single point to
a trajectory with some degree of spatial resolution. However,
they are significantly influenced by interference factors. For
example, vehicular trajectories are affected by the driving
state and road conditions. In contrast, continuous monitoring
platforms typically offer lower spatial resolution, but provide
stable and long-term monitoring capabilities.
Ground-based research stations, as examples of continuous
observation platforms, are a vital component of atmospheric (a) Aerosol, Clouds and Trace gases Research Infrastructure
Stations (ACTRIS) and NOAA Earth System Research Labora-
research infrastructure. For instance, the SMEAR (Stations tories (ESRL) networks
for Measuring Ecosystem–Atmosphere Relations) stations in
Finland provide continuous, long-term measurements of atmo-
spheric, ecological and meteorological variables [20]. These
stations generate critical data on greenhouse gas concentra-
tions, aerosols, and other atmospheric constituents [23], which
are instrumental, e.g., for calibrating AI models for air quality
sensor calibration and developing virtual sensors [24].
Research infrastructure networks such as ACTRIS (Aerosol,
Clouds and Trace Gases Research Infrastructure) [25] and
ICOS (Integrated Carbon Observation System) [26] provide
coordinated and standardized data from a wide range of
monitoring stations across Europe. Another example is the
Global Atmosphere Watch (GAW) program, coordinated by (b) Integrated Carbon Observation System (ICOS) networks
the World Meteorological Organization (WMO), which fo-
cuses on building a coordinated global understanding of atmo-
spheric composition and its changes and on helping to improve
the understanding of interactions between the atmosphere, the
oceans, and the biosphere. As shown in Figures 2a, 2b, and
2c, the ACTRIS, ICOS, and GAW networks span diverse geo-
graphical locations. These networks incorporate advanced and
high-quality instrumentation, ensuring consistent collection of
reliable and standardized data. This consistency is vital for AI-
driven atmospheric research, supporting accurate and scalable
atmospheric observations and simulations.
Remote sensing approaches, in contrast, often rely on opti-
cal sensing technologies to perform atmospheric observations. (c) Global Atmosphere Watch (GAW) Global stations
These approaches can be divided into air- and space-based Fig. 2: Examples of the locations of atmospheric observation
observation platforms. Air-based platforms, which include air- infrastructure, with station data retrieved in 2024. These net-
craft, tethered balloons, and unmanned aerial vehicles (UAVs), works operate continuously to monitor a wide range of atmo-
can achieve large-scale observations by mounting instruments spheric variables, providing essential data for environmental
on these platforms. The ability to control flying altitude research, modeling, and policy support.
enables the retrieval of vertical atmospheric profiles. How-
ever, air-based platforms are also restricted by meteorological
conditions. Space-based platforms, such as satellites, have the and air quality [31]. These systems complement traditional
advantage of a broad field of view, enabling the observation approaches by filling observation gaps and providing data for
of large areas or even the entire globe in a short period. AI-based analytics and decision-making [32]. At the same
However, spatial resolution varies widely between satellites, time, they also face challenges related to security, privacy, and
some capable of sub-meter resolution and others only at the the need for robust connectivity.
kilometer scale [27]. Nevertheless, satellite observations are One major challenge in atmospheric science is the in-
indispensable for understanding the Earth’s atmosphere [28]. tegration of data from different sources and the need for
Emerging technologies such as the Internet of Things (IoT) consistent data standards across various research networks and
offer effective solutions for atmospheric observations [29]. stations [33]. Ensuring data interoperability is crucial for AI
IoT-based systems integrate a network of low-cost, distributed models, which often require large and diverse datasets for
sensors that provide real-time, high-resolution data at local training. Inconsistent data formats, measurement techniques,
scales, such as for monitoring climate [11], weather [30] and reporting standards across different research networks
 4

Fig. 3: Locations of selected High-Performance Computing (HPC) facilities dedicated to weather forecasting (blue) and weather
and climate research (red), based on data retrieved from the TOP500 list in 2024.

can lead to challenges in data harmonization and limit the researchers to handle, process, and analyze large-scale, multi-
effectiveness of AI applications [34]. dimensional datasets critical for advancing our understanding
Another significant challenge pertains to the operation and of atmospheric systems. As shown in Figure 1 (the upper
maintenance of an extensive environmental monitoring infras- part), these platforms encompass high-performance computing
tructure. Networks of ground-based stations, remote sensing (HPC), cloud computing, fog computing, and edge computing.
equipment, and IoT sensors require regular calibration, data Each serves a unique role in data processing and analytics,
synchronization, and fault management to ensure accuracy catering to different scales, latency requirements, and compu-
and reliability [12]. Maintaining this massive infrastructure tational needs.
not only demands considerable resources but also introduces High Performance Computing (HPC) platforms provide
logistical challenges, as sensors and stations deployed in re- centralized, large-scale computational resources optimized
mote or extreme environments often face accelerated wear and for performing highly intensive tasks [35], such as climate
calibration drift, impacting data quality and consistency [13]. modeling [36], real-time weather prediction [37], and atmo-
Concurrently, modern data science approaches, particularly spheric chemistry simulations [38]. HPC services are typically
those based on AI and machine learning (ML), enable re- provided by national governments, universities, and private
searchers not only to handle massive volumes of data but also institutions to support scientific research and development.
to effectively manage and optimize large-scale atmospheric These systems consist of powerful supercomputers designed
infrastructure [15]. Through AI-driven solutions, atmospheric to handle complex calculations and process vast datasets in
scientists can automate critical tasks such as sensor calibration, parallel efficiently. Figure 3 shows the top 500 HPC dedicated
anomaly detection, and system maintenance across extensive to weather and climate research activities. One prominent
measurement networks, ensuring high data quality and reli- example is the European Centre for Medium-Range Weather
ability of distributed monitoring stations while significantly Forecasts (ECMWF), which leverages HPC infrastructure to
reducing operational overhead [16]. By integrating advanced perform global weather forecasts that demand high computa-
computing infrastructure with AI, atmospheric science can tional precision and rapid processing [39]. HPC platforms are
gain deeper insight into atmospheric processes and envi- particularly well-suited for tasks requiring maximum process-
ronmental changes, maintaining rigorous standards of data ing power and minimal latency, making them indispensable
accuracy and enhancing infrastructure efficiency. for advancing atmospheric sciences. However, their reliance on
centralized infrastructure and robust connectivity poses certain
limitations in accessibility and flexibility.
B. Computational and Analytical Platforms Cloud Computing offers flexible, on-demand access to
Computational and analytical platforms are foundational computational resources and tools via remote servers, making
components of atmospheric science infrastructure, enabling it highly scalable and accessible [40]. Unlike HPC, which
 5

is typically government-provided, cloud services are usually infrastructure still hinder these researchers, again particularly
offered by private companies and primarily rely on large-scale the developing countries that are the most vulnerable [51]. In
data centers. However, they can also utilize distributed archi- addition, maintaining and upgrading existing research stations
tectures that operate across multiple geographic regions [41]. and infrastructure networks requires significant investment and
Cloud platforms are increasingly essential for handling large- coordination at the international level [52].
scale computational demands of climate and environmental Advancements in research infrastructure are underway to
research, offering scalability and flexibility to meet diverse address these challenges. Efforts to standardize data formats
research needs. They support applications such as distributed and improve data sharing among international networks, such
analysis of atmospheric datasets, integration of observational as the WMO Integrated Global Observing System (WIGOS),
and model data, and even AI-driven processes such as machine are helping to create a more unified framework for atmospheric
learning model training and predictive analytics [42]. For research [53], [54]. Furthermore, the development of cloud-
example, Google has developed a cloud computing platform, based platforms such as Google Earth Engine [43] and the
called Google Earth Engine (GEE), to effectively facilitate Copernicus Climate Data Store [55] provides easier access to
processing large geo data over large areas and monitoring the vast datasets and computational resources, allowing a broader
environment for long periods of time [43]. Furthermore, by range of researchers to engage in AI-driven atmospheric
leveraging adaptable on-demand resources, cloud platforms studies.
reduce the overhead of maintaining physical infrastructure, To summarize, a robust and integrated research infras-
making them particularly advantageous for interdisciplinary tructure, both for collecting observations and analyzing the
research teams and projects requiring dynamic resource allo- measurements, is critical for advancing the state of atmo-
cation [44]. spheric research. Without sufficient data, the potential benefits
Fog Computing serves as an intermediary layer that extends of AI models are hard to achieve, whereas computational
cloud services closer to local edge devices, enabling faster data infrastructure is needed to harness these models. There are
processing and reduced latency for time-sensitive applications. many kind of systems for both purposes, ranging from HPC
This proximity reduces latency, improves efficiency, and en- systems, cloud platforms, research stations, and international
hances performance for time-sensitive applications [45]. For monitoring networks, with all of them having their own
example, a service framework employing a distributed model advantages and disadvantages. To this end, it is essential to
averaging algorithm based on the fog computing paradigm and understand how these infrastructure can be best harnessed
federated learning is proposed to facilitate accurate data fusion to advance the adoption of AI applications in atmospheric
and real-time analysis of heterogeneous air quality data in sciences. Continued efforts to improve data interoperability
Beijing, China. This approach successfully integrated diverse and broaden access to computational resources will further
data sources to deliver high-precision insights for air quality empower researchers to address complex challenges related to
monitoring [46]. climate change, air quality, and atmospheric dynamics. Table I
Edge Computing takes computation even closer to the data provides a summary of the state-of-the-art infrastructure, key
source than fog computing, typically directly on the sensing challenges, and potential AI-driven solutions for both cate-
device or the nearby gateway. It enables ultra-low latency gories of infrastructure.
and minimal reliance on connectivity by processing data
locally [47]. Edge platforms are best suited for applications III. A RTIFICIAL I NTELLIGENCE FOR ATMOSPHERIC DATA
where immediate local data processing is required to reduce S CIENCES
latency. This also holds for many AI-driven environmental The infrastructure introduced in the previous section pro-
monitoring systems, such as detecting harmful gases [48] or vides the foundation for leveraging AI by enabling both the
fire hazards [49]. Another example is the integration of AI collection of relevant data and the hardware on which to
algorithms, including deep learning techniques, to process model and analyze it. In this section, we shift our focus to
hyperspectral imaging data on FPGA edge devices installed the AI models themselves, exploring various examples of how
onboard unmanned aerial vehicles (UAVs). These UAVs pro- AI is being utilized across different domains of atmospheric
vide a flexible and mobile solution for air quality monitoring, sciences. We discuss specific applications, key challenges, and
enhancing spatial coverage and complementing continuous recent advancements, highlighting the transformative potential
monitoring stations by accessing areas that are difficult to of AI in addressing contemporary environmental challenges.
reach with continuous monitoring stations [50].

A. Overview of Artificial Intelligence

C. Challenges and Possible Solutions This section provides an overview of Artificial Intelligence
Despite the many strengths of the current research infras- (AI) and its subsets, highlighting their tasks and methods
tructure, there are also challenges. One of the most sig- relevant to applications in atmospheric sciences.
nificant challenges is the uneven accessibility of computa- 1) AI methodologies: Artificial Intelligence (AI) encom-
tional resources, particularly for researchers in developing passes a broad range of technologies and methodologies that
countries or institutions with limited access to HPC systems. enable machines to simulate human intelligence processes,
Although cloud infrastructure offers potential for overcoming including learning, reasoning, and decision-making. In atmo-
accessibility barriers, disparities in internet connectivity and spheric and environmental monitoring, AI holds significant
 6

TABLE I: Summary of infrastructure for atmospheric monitoring and analytics

Infrastructure State-of-the-Art Challenges Potential Solutions
Observation and Sens- A combination of mobile, continuous, Ensuring continuous operation and AI-driven automation for sensor cali-
ing Systems airborne, space-based, and IoT-enabled maintenance of extensive monitoring bration, anomaly detection, and predic-
observation networks networks, including calibration, data tive maintenance can optimize infras-
synchronization, and fault management tructure performance and reduce oper-
to maintain accuracy and reliability ational costs
Computational and A diverse computing ecosystem, includ- Data interoperability issues due to vary- International organizations like WMO
Analytical Platforms ing HPC, cloud, fog, and edge comput- ing formats across research networks, and ACTRIS should standardize data
ing platforms along with limited HPC access in certain formats and promote data sharing.
regions Cloud-based platforms such as Google
Earth Engine and Copernicus Climate
Data Store enhance accessibility to large
datasets and computational resources,
enabling broader AI-driven research

potential to revolutionize the observation, modeling, and pre- layered neural networks, gained prominence due to its abil-
diction of atmospheric phenomena. As the field of atmospheric ity to process high-dimensional, non-linear data. DL models
science has transitioned into the era of big data, AI has become have significantly enhanced atmospheric applications, such as
indispensable for uncovering patterns, extracting insights, and improving the detection and prediction of extreme weather
automating decision-making processes. events [58] and automating the classification of satellite cloud
imagery with high accuracy and efficiency [59].
Building upon these developments, the field is now moving
toward generative AI, which includes Foundation Models and
Large Language Models (LLMs) [60]. These models not only
learn from massive multimodal datasets but can also gener-
ate new data, assist in hypothesis generation, and simulate
plausible environmental scenarios. Generative AI holds par-
ticular promise for atmospheric sciences in tasks such as data
synthesis, knowledge extraction from heterogeneous sources,
and support for decision-making under uncertainty. Figure 4
illustrates the hierarchical relationship among various AI do-
mains and highlights some widely used techniques, including
ML and DL approaches relevant to atmospheric sciences. As
these technologies continue to evolve, they promise to further
enhance our understanding of atmospheric phenomena and
improve our ability to respond to environmental challenges.
2) Machine Learning: Machine learning forms the core of
most AI applications in atmospheric sciences, enabling the
analysis of complex datasets and the extraction of valuable
insights, making it essential for advancing research and op-
erational applications. ML can be broadly classified into four
categories: supervised learning, unsupervised learning, semi-
Fig. 4: Artificial Intelligence with its prominent subsets supervised learning, and reinforcement learning.
Supervised learning relies on labeled data, where models
are trained using input-output pairs to make predictions or
Among the various AI methodologies, Machine Learning classifications. The availability of high-quality labeled data
(ML) serves as a key subset that focuses on algorithms is crucial for the success of these models, as it directly
enabling systems to learn from data and make data-driven affects their performance and accuracy [61]. Algorithms such
predictions. Initially, the integration of AI in atmospheric sci- as support vector machines and neural networks have been
ences began with ML techniques, which provided data-driven extensively applied in atmospheric sciences for tasks such as
alternatives to traditional numerical modeling approaches. ML weather forecasting [62] and classifying sources influencing
algorithms are designed to learn from historical datasets and air quality indices [63].
make predictions or classifications without explicitly pro- Unsupervised learning, in contrast, deals with unlabeled
grammed instructions. These methods have proven particu- data, identifying hidden structures or patterns within datasets.
larly effective for applications such as air quality forecasting, Methods such as k-means and hierarchical clustering are
anomaly detection, and climate trend analysis [56], [57]. commonly used in atmospheric science to analyze air pol-
As the volume and complexity of data grew, Deep Learning lutants, including identifying pollution sources, investigating
(DL), a specialized area within ML that employs multi- long-range pollutant transport pathways, and supporting the
 7

development of effective mitigation strategies [64]. simulations and faster analyses, highlighting the transformative
Semi-supervised learning bridges the gap between super- potential of AI in atmospheric research.
vised and unsupervised methods by leveraging a small amount ML and DL techniques have also been employed to predict
of labeled data along with a larger pool of unlabeled data. air quality levels with higher accuracy and efficiency than
This approach is particularly valuable in atmospheric appli- traditional models can achieve [74]. For example, super-
cations where labeled data is scarce or expensive to obtain, vised learning algorithms such as support vector machines
such as detecting extreme weather events using multichannel (SVMs) [75] and random forests [76] have been used to
spatiotemporal convolutional neural networks [58]. forecast pollutant concentrations based on historical data and
Reinforcement learning, a goal-oriented learning paradigm meteorological variables. Similarly, deep learning models such
where an agent interacts with its environment and learns as convolutional neural networks (CNNs) [77] and long short-
through feedback in the form of rewards or penalties, has term memory (LSTM) networks [78] have demonstrated suc-
thus far been largely underexplored in atmospheric sciences cess in capturing complex, non-linear relationships between
due to challenges in creating realistic, interactive environments pollutants and environmental factors. Semi-supervised learning
and the computational demands of simulations. However, has been employed to achieve fine-grained air quality fore-
reinforcement learning holds promise for enabling automated casting. For example, self-supervised hierarchical graph neural
decision-making and actions derived from atmospheric moni- networks can enhance predictive accuracy even in scenarios
toring and modeling, a topic we return to in Section IV. with limited labeled data, thereby expanding the applicability
Figure 5 illustrates the applications of different learning of AI-driven models in diverse monitoring environments [79].
methods in atmospheric sciences and the following subsections The trend to improve air quality forecasting systems has
explore their applications across different fields within atmo- led to the development of new methods that utilize modern
spheric sciences. Understanding these various ML method- observational data in models, including data assimilation and
ologies is crucial for harnessing their potential to address ML methods that combine and fuse information from multiple
complex challenges in atmospheric research and operational sources. These sources include low-cost sensor and medium-
applications, and to identify the future research directions. cost sensor networks, permanent monitoring networks, and
UAV-based, aircraft-based, and satellite-based measurements
(both in situ and remote sensing) [69]. This multi-source
B. Air Quality Monitoring and Modeling data fusion enhances the spatial and temporal resolution of air
Air quality monitoring and modeling refer to the processes quality predictions and fosters synergies between atmospheric
and techniques used to observe, analyze, and predict the sciences and computational technologies. Other AI techniques
concentrations of air pollutants in the atmosphere. Common such as clustering and anomaly detection have also proved
pollutants include particulate matter (PM), carbon monoxide useful, e.g, to identify pollution sources [80], track their
(CO), nitrogen oxides (NOx), sulphur dioxide (SO2), ozone dispersion [81], and even detect unusual pollution events [82].
(O3) and particle number concentration (PNC). These pollu- In urban environments, where industrial activities and traffic
tants have significant implications for public health [65], [66], are major contributors to air pollution, these AI techniques
ecosystems [67], and climate [68]. have been particularly useful in understanding the impact of
Monitoring involves the use of various sensors and instru- urbanization and industry on atmospheric composition.
ments, such as those described in the previous section, placed Despite these advancements, several challenges remain in
at ground-based stations or integrated into satellite systems the integration of AI with air quality monitoring. Data scarcity
to collect real-time data on air pollutants [27]. In contrast, and data quality issues can introduce biases into AI mod-
modeling employs mathematical and computational methods els [83]. Common concerns include noise and gaps in sensor
to simulate air quality scenarios, identify pollution sources, measurements. The interpretability of AI is another significant
predict future pollutant levels, and assess the impact of various concern as many of the models operate as ”black boxes”
factors such as weather patterns, urbanization, and industrial that offer limited information into the underlying physical
activities [69]. The most widely adopted practice for air qual- processes governing air pollution [84] or the limitations of
ity observation have relied on ground-based monitoring sta- sensor technology [85]. Another challenge is the need for
tions [70] and chemical transport models (CTMs) to simulate AI models to generalize well across different geographic
the dispersion and chemical transformation of pollutants [71]. regions and climatic conditions, as pollutant behavior can vary
However, these methods face challenges, including limited significantly [86], [61].
spatial resolution [24] and high computational costs [13]. One of the key future directions for air quality monitoring
In recent years, the integration of AI into air quality and modeling is the development of hybrid models that com-
monitoring and modeling has transformed the field. AI-based bine the strengths of AI and CTMs. These hybrid models aim
models have been used to emulate the outcomes of extensive to leverage the data-driven accuracy of AI while maintaining
CTM numerical calculations at significantly lower computa- the interpretability and robustness of physical models [87].
tional costs [72]. For instance, deep learning methods have This approach exemplifies how bridging AI and atmospheric
been employed to emulate a typical gas-phase chemistry solver sciences can enhance both fields, leading to a deeper under-
implemented in CTMs, achieving computational efficiency standing of atmospheric phenomena and improved predictive
improvements of 10.6 times on one CPU and 85.2 times on capabilities. Another important direction is the deployment of
one GPU [73]. These substantial gains enable more extensive Internet of Things (IoT) devices with monitoring capabilities,
 8

1‌ Emulator for‌
Detecting‌‌
3‌ ‌c onventional‌ ‌
unlabbeled‌ sub-grid
parameterization‌ ‌ 2‌
‌Cloud images‌‌ in‌ ‌
Hyperspectral image global model‌ Studying‌‌
atmospheric‌‌
2‌ classification for
environmental
dynamic‌ 3‌
assesments
Earth System
Satellite‌ Modeling and‌
Cloud images‌‌ Comparison‌
Remote‌ ‌ Climate Change‌
Classification‌ Different‌
Sensing‌
1‌ ‌Climate Models‌

Artificial‌‌ 1‌
Clustered ML
models for solar
Intelligence‌‌ CTM
irradiance for‌ Emulator‌
forecasting‌ Air Quality‌‌
3‌ Operational‌ Atmospheric‌ Monitoring
Describing‌
Meteorology‌ Advertising‌
Sciences‌
Screening‌
and‌‌ 2‌
Interviewing‌ Onboarding‌
Detection of Modeling‌ Fine -grain
extreme‌‌ AQ‌
weather event‌ ‌forecasting‌

2‌ In-situ‌‌

Weather forecast‌
Atmospheric‌ Identify‌ 3‌
monitoring‌ ‌ ollution
p
sources‌

1‌ Undestanding‌ Integrated‌ NPF‌‌

key drivers to‌ aerosol data‌ Identification‌ 1‌
NPF event‌ ‌of satelite‌‌ 1‌ ‌=‌ ‌Supervised Learning‌
3‌
and in -situ‌‌
2‌
2 =‌ ‌Semi-supervised Learning‌
measurement‌
3 =‌ ‌Unsupervised Learning‌

Fig. 5: The applications of supervised, semi-supervised and unsupervised learning methods in different fields of atmospheric
sciences.

which is expected to significantly enhance the field by pro- several atmospheric research stations that are used to perform
viding near-real-time predictions and actionable insights for extensive measurements of atmospheric concentrations.
decision-makers [16]. These efforts are critical for developing Historically, data analysis in atmospheric science has relied
more effective environmental monitoring systems and boosting on statistical methods. However, the increasing volume and
disaster prediction and climate change mitigation efforts. complexity of modern datasets demand advanced AI methods
to process and analyze the vast quantities of data ground-
based measurement stations generate [19]. ML techniques
C. In-situ Atmospheric Monitoring have emerged as valuable tools, capable of automating data
In-situ atmospheric monitoring involves to the direct, con- analysis, detecting trends, identifying anomalies, and predict-
tinuous observation and collection of data on various atmo- ing future atmospheric events [18]. For instance, ML and DL
spheric variables at ground-based monitoring stations. Unlike models have been successfully applied to study atmospheric
traditional air quality monitoring, which primarily targets new-particle formation (NPF), a key process in generating
pollutants in urban areas, in-situ atmospheric monitoring tracks climatically significant aerosol particles. NPF events, detected
a wider range of atmospheric variables, such as aerosols, through changes in aerosol particle size distributions, tradi-
greenhouse gases (e.g., CO2 and CH4) and other trace gases tionally require manual classification of measurement days
[88]. The primary goal is to provide high-resolution, real- which is time-intensive and prone to human bias [89], [90],
time data that accurately reflect the atmospheric composition [91]. Recently, automated ML-based classification of NPF
at specific locations, aiding in climate modeling and weather events has significantly reduced manual effort and enabled the
prediction. This detailed data serves as a crucial input for un- rapid creation of event datasets, facilitating further analysis by
derstanding atmospheric processes, validating remote sensing atmospheric scientists [92], [93].
observations, and improving the precision of climate change Unsupervised learning can also be highlight helpful at
projections [22]. As an example initiative, the Aerosol, Clouds, automating manual analysis processes. For example, mu-
and Trace Gases Research Infrastructure (ACTRIS) is a pan- tual information [94] has been used to quantify non-linear
European research infrastructure that produces high-quality relationships between various measured variables and NPF
data and information on short-lived atmospheric constituents events [95], [96]. Methods such as these combined with big
and the processes driving their variability in both natural data analytics frameworks enable scientists to mine these rela-
and controlled atmospheres. ACTRIS stations are typically tionships, uncover associations, and generate new hypotheses
equipped with particle number counters to monitor atmo- that might otherwise remain hidden. Semi-supervised learning
spheric composition. Figure 2a illustrates the locations of methods can also reduce manual effort, particularly in cases
 9

where large volumes of unlabelled data are supplemented by weather radars, satellite remote sensing, and meteorological
a limited number of labeled samples. For example, semi- models [105]. AI models excel at handling large datasets, iden-
supervised learning methods have been used to integrate tifying complex patterns, and making data-driven predictions
aerosol data from multiple satellite instruments alongside high- that enhance traditional weather forecasting methods [102].
accuracy, ground-based instruments, enhancing the reliability For example, convolutional neural networks (CNNs) have been
and scalability of aerosol predictions [97]. applied to satellite images to improve cloud classification
Despite these advancements, significant challenges remain and storm tracking [106], [107]. Similarly, recurrent neural
in applying AI to in-situ atmospheric measurement data. networks (RNNs), especially long short-term memory (LSTM)
Similar to air quality modeling, the black-box nature of ML networks, have been utilized to model temporal dependencies
models is a major concern, as it complicates the interpre- in weather data, leading to more accurate short-term weather
tation of their decision-making processes and obscures the predictions [108].
underlying atmospheric dynamics. This lack of transparency One of the most transformative applications of AI in opera-
limits trust and hinders the adoption of AI-driven insights tional meteorology is enhancing ensemble weather forecasting,
in critical applications. Moreover, current data science meth- which traditionally involves running multiple simulations with
ods often overlook the importance of learning cause-and- slight variations in initial conditions to account for atmo-
effect relationships from observations, focusing instead on spheric uncertainties [109]. AI-driven models, particularly
correlations without considering the physical and generative generative adversarial networks (GANs), have been integrated
processes that govern atmospheric systems. Such limitations into the post-processing of ensemble forecasts, significantly
reduce the generalizability and scientific interpretability of AI- reducing error margins and increasing the reliability of prob-
based approaches [15]. abilistic predictions [110]. In addition, AI has improved now-
To address these challenges, future research must prioritize casting, the short-term forecasting of weather over the next few
the development of AI models that are both interpretable hours, by processing real-time radar and sensor data to deliver
and trustworthy. Integrating explainable AI (XAI) techniques accurate predictions on rapid timescales, which is essential for
can provide the necessary transparency in decision-making, severe weather warnings [111].
enabling researchers to trace model outputs back to their Semi-supervised learning techniques, such as a multichannel
contributing factors [98]. This enhances model credibility spatiotemporal CNN architectures, have advanced the detec-
and fosters confidence in AI-driven atmospheric predictions. tion of extreme weather events such as hurricanes, extra-
Furthermore, incorporating causal inference methods into AI tropical cyclones, and weather fronts, particularly when avail-
frameworks is essential to ensure that models learn meaningful able data lacks comprehensive labeling [58]. Unsupervised
cause-and-effect relationships rather than relying solely on learning has also been applied to short-term solar irradiance
statistical correlations [99]. Such advancements are crucial forecasting. Forecasts are generated by using the most suitable
for enhancing our understanding of atmospheric dynamics. In models in different clusters that help to improve the accuracy
addition, developing interactive AI systems will allow human of one-hour-ahead forecasts for global horizontal irradiance
researchers to query and scrutinize the internal reasoning of (GHI), a key parameter for solar energy management [112].
models, facilitating collaboration between domain experts and These AI applications illustrate the growing role of machine
AI-driven analytics [18]. learning techniques across varied facets of operational meteo-
rology, from managing uncertainties in ensemble forecasting to
making real-time predictions and optimizing solar forecasting.
D. Operational Meteorology Despite these advancements, challenges persist in the appli-
Operational meteorology refers to the application of me- cation of AI to operational meteorology. As before, the limited
teorological science in real-time settings to support decision- interpretability of AI models is a critical concern, raising
making, forecasting, and public safety [100]. This involves the concerns for operational meteorologists who need to trust and
use of observational data, numerical weather prediction (NWP) understand the models they rely on for decision-making [113].
models, and forecasting tools to produce weather forecasts Another challenge stems from the inherent unpredictability
and warnings that are disseminated to governments, industries, of the atmosphere. Specifically, while AI models excel at
and the public. Accurate weather forecasting plays a critical capturing patterns from historical data, they often struggle to
role in agriculture, aviation, disaster preparedness, and daily generalize to novel atmospheric conditions, particularly in the
life, making it an essential part of modern society [101]. face of climate change and extreme weather events [114]. Fur-
Standard practice in operational meteorology involves using thermore, integrating AI with traditional NWP models requires
NWP models that employ mathematical equations to simulate substantial computational resources, making it challenging
atmospheric processes [102], [103]. However, the advent of to balance the strengths of physics-based models with the
high-performance computing (HPC), satellite observations, efficiency of data-driven approaches [115].
and the Internet of Things, has lead to an explosion in the To address these challenges, future research should focus on
volume of atmospheric data available for analysis, presenting developing hybrid models that combine the physical insights
both challenges and opportunities for the field [104]. of NWP with the predictive capabilities of AI. These models
AI techniques, particularly ML and DL, have revolutionized aim to leverage the strengths of both approaches, enhanc-
weather forecasting by enabling the analysis of vast amounts ing forecast accuracy while preserving the interpretability
of data from diverse sources, including ground-based stations, and reliability required for operational meteorology [115].
 10

AI can also be integrated into specific tasks such as bias

correction, downscaling, and data assimilation, ensuring that
physics-based constraints remain central to the forecasting
process. In addition, advancements in edge computing and IoT
technologies could enable real-time data collection and AI-
driven analysis, allowing for more precise localized weather
predictions with reduced reliance on centralized computational
resources [30]. These innovations will facilitate more efficient
and scalable meteorological models, bridging the gap between
AI and traditional forecasting techniques.

E. Satellite Remote Sensing

Remote sensing involves acquiring data about the Earth’s
surface and atmosphere using sensors mounted on satellites Fig. 6: Accumulated number of Earth observation satellites
or airborne platforms. These sensors detect electromagnetic (since 1995)
radiation that is either reflected or emitted from the Earth,
which is subsequently processed into images or datasets that
reveal critical environmental parameters [116]. Satellite remote utilized for tasks like hyperspectral image (HSI) classification,
sensing plays a vital role in atmospheric sciences, particularly where limited labeled data restricts traditional training. These
in addressing global-scale monitoring gaps that cannot be techniques enable enhanced feature extraction and classifi-
covered by ground-based sensors alone. For example, over cation for detailed environmental assessments [125], [126].
oceans, which constitute more than 70% of the Earth’s surface, Unsupervised learning methods are increasingly critical in
there is a lack of fixed air quality towers or IoT sensors, remote sensing, particularly in point cloud analysis, where
making satellite observations indispensable. Remote sensing the absence of labeled data limits traditional approaches. For
enables the measurement of atmospheric pollutants such as example, unsupervised point cloud detection has emerged as
aerosols, nitrogen dioxide (NO2), and ozone (O3) over both a key technique, enabling generalized representations from
land and sea, supporting air quality assessments even in unlabeled data that effectively support diverse applications in
remote or data-sparse regions [117]. In addition, satellite-based environmental monitoring [127].
measurements are critical for understanding ocean-atmosphere One key advancement enabled by AI is the improvement
interactions, including ocean acidification and carbon fluxes. in the spatial and temporal resolution of satellite-derived
These systems help quantify oceanic uptake of CO2, shedding products. Traditional remote sensing systems often face lim-
light on the ocean’s role as a carbon sink and its influence on itations in their ability to provide high-resolution data due
global climate dynamics [118]. to trade-offs between spatial coverage and frequency of ob-
Figure 6 illustrates the accumulated number of Earth obser- servations. AI techniques, such as super-resolution models,
vation satellites (since 1995). Over the past three decades, the have been developed to enhance the resolution of satellite
number of satellites dedicated to Earth observation has surged, images, allowing for more detailed monitoring of small-scale
significantly enhancing the capability to monitor various as- environmental changes [128], [129]. Furthermore, AI-based
pects of the Earth’s atmosphere, oceans, land surfaces, and algorithms have been employed to fuse data from multiple
vegetation. Satellite remote sensing has become indispensable satellites and sensors, creating comprehensive datasets that
for climate change research, disaster management, agriculture, offer more accurate and holistic insights into the Earth’s
and weather forecasting [119], [120]. atmospheric and environmental systems [130], [131].
In recent years, AI has played an increasing role in en- Despite these advancements, several challenges remain in
hancing the analysis of satellite remote sensing data, allow- the application of AI to satellite remote sensing. The primary
ing for more detailed environmental monitoring and assess- challenge is the vast volume and diversity of data generated
ment [121]. By applying AI to extract meaningful features by multiple satellites, which requires powerful computational
from vast, complex datasets, researchers can improve the resources and efficient algorithms to process [132]. Further-
spatial and temporal resolution of remote sensing applications, more, AI models used for remote sensing data analysis often
which results in more accurate insights into environmental require large, annotated training datasets, which are not always
changes [122]. For instance, convolutional neural networks readily available [133]. Finally, as with the other fields, also
(CNNs) are widely used for the automated classification and remote sensing would require for explainable models that
recognition of satellite cloud images, as demonstrated on allow interpreting the physical meaning of their outputs [134].
large-scale cloud image databases [59]. In tandem, recurrent To overcome these challenges, future research should focus
neural networks (RNNs) and long short-term memory (LSTM) on the development of explainable AI-based models and
networks analyze temporal sequences within satellite data, hybrid models that combine the strengths of both AI and
enabling better predictions of atmospheric variables and the traditional remote sensing techniques [121]. In addition, efforts
monitoring of dynamic events like forest fires and deforesta- to create more comprehensive, annotated datasets will be
tion [123], [124]. Semi-supervised learning techniques are crucial for improving the performance of AI models in satellite
 11

data analysis [59], [133]. As computational power continues methods, such as variational autoencoders (VAEs), offer fur-
to grow and more advanced AI algorithms are developed, the ther advancements by performing dimensionality reduction
integration of AI with satellite remote sensing will likely yield and density estimation on climate data. VAEs automatically
even greater improvements in environmental monitoring and derive low-dimensional, physically meaningful representations
climate science. of complex datasets from various climate models, uncovering
new notions of similarity and variability in atmospheric pat-
terns [144].
F. Earth System Modeling and Climate Change
In addition to improving predictive accuracy, interpretability
Earth system modeling (ESM) is a multidisciplinary ap- and understanding are essential for the effective use of DL in
proach to simulating the complex interactions between the Earth system models. Interpretability remains a challenge for
Earth’s atmosphere, oceans, land, and biosphere. These models deep neural networks, as they are often not self-explanatory
integrate various physical, chemical, and biological processes and may struggle to identify causal relationships from obser-
to study the behavior of the Earth as a single, interconnected vational data. Even modern Earth system models suffer from
system [135]. One of the primary objectives of ESM is to limited traceability to their assumptions, which reduces their
improve our understanding of climate change by simulating interpretability. Ensuring physical consistency in predictions is
future climate scenarios based on different greenhouse gas another important challenge. Specifically, while deep learning
emission trajectories [136]. Climate change, driven by human models fit data well, they may produce physically implausible
activities such as industrialization and deforestation, has led outcomes due to observational biases or extrapolation. Inte-
to significant changes in the Earth’s climate, including rising grating domain knowledge and physical laws into models can
temperatures, shifting weather patterns, and an increasing address this issue. Furthermore, DL must effectively handle
frequency of extreme weather events [137]. ESMs are critical complex, noisy, and high-dimensional data, necessitating new
for predicting how these changes will evolve over the coming network architectures that capture both local and long-range
decades and centuries, providing essential information for relationships. Managing uncertainty will require integrating
policymakers and scientists working on climate mitigation and Bayesian or probabilistic methods. In addition, many DL
adaptation strategies [138]. applications face the challenge of limited labeled data, high-
DL models have recently emerged as powerful tools for lighting the need for methods that leverage unlabeled data
improving the accuracy and efficiency of Earth system model- through unsupervised or semi-supervised learning. Finally, the
ing, particularly in climate prediction and the study of climate computational demands of geoscience problems are immense,
change [139]. Traditional ESMs are computationally intensive as demonstrated by platforms like Google’s Earth Engine,
and rely heavily on physics-based equations to simulate the which can help manage these challenges and may lead to
Earth’s climate. Although these models provide detailed in- further deep learning applications [140].
sights into physical processes, they often struggle with high In terms of advancements, ongoing research aims to improve
computational costs and long simulation times. In contrast, DL the scalability of DL models for global-scale climate simu-
models, such as CNNs and RNNs, have shown promise in cap- lations, leveraging high-performance computing (HPC) and
turing complex, non-linear patterns in climate data, enabling cloud-based platforms [145], [146]. Furthermore, efforts are
faster and more efficient simulations [140]. For instance, DL being made to develop more interpretable AI models, such as
techniques have been applied to represent subgrid atmospheric physics-informed neural networks (PINNs), which incorporate
processes in climate models, learning directly from multiscale physical laws into DL architectures to enhance both accuracy
models that explicitly resolve convection. These trained DL and transparency [147], [148].
models can replace conventional subgrid parameterizations
in global general circulation models, allowing them to in-
teract dynamically with resolved atmospheric processes and G. Other Environmental Challenges and Atmospheric Science
surface-flux schemes. This approach significantly enhances Research
the representation of complex, nonlinear subgrid processes, Atmospheric science intersects with various environmental
particularly cloud dynamics, which have historically been a disciplines, highlighting the broader impact of AI applica-
significant source of uncertainty in climate models [141]. tions beyond traditional meteorological studies. AI techniques
Another example of DL in ESMs is its use in forecasting the developed for atmospheric research can be extended to ad-
El Niño/Southern Oscillation (ENSO); a DL-based statistical dress other environmental challenges, contributing to a more
model has shown high accuracy in predicting ENSO patterns, comprehensive understanding of Earth system processes. For
which is crucial for anticipating extreme regional climate example, the coupled interactions between the atmosphere,
events and their ecosystem impacts [142]. ocean, and land surface form the backbone of Earth system
To address the common challenge of limited labeled data, dynamics, influencing climate variability, weather extremes,
semi-supervised learning (SSL) methods are also being devel- and long-term environmental trends [149]. AI is increasingly
oped for diverse atmospheric applications. SSL is particularly used to capture these cross-domain processes by analyz-
valuable in studying atmospheric dynamics, where large-scale ing multivariate datasets from satellites, in-situ sensors, and
labeled datasets are sparse, enabling models to leverage both reanalysis products. For example, ML models help reveal
labeled and unlabeled data to enhance predictive performance how oceanic conditions such as sea surface temperature and
across atmospheric variables [143]. Unsupervised learning salinity influence atmospheric circulation, while also assessing
 12

TABLE II: Summary of AI applications in atmospheric sciences

Field State-of-the-Art Challenges Possible Solutions
Air Quality Monitor- AI enhances air quality predictions, Data scarcity, model interpretability, and Hybrid AI-physics models, expanded
ing and Modeling source identification, and pollutant dis- generalization across regions with di- sensor networks, and real-time IoT-
persion modeling. ML and DL methods verse pollution dynamics. based monitoring for improved accu-
improve computational efficiency over racy.
chemical transport models (CTMs) and
detect pollution anomalies.
In-situ Atmospheric AI automates data analysis, trend de- ML models often act as black boxes, Explainable AI (XAI) for model inter-
Monitoring tection, and anomaly identification in limiting transparency. Causal relation- pretability, causal inference techniques,
ground-based observations, improving ships in atmospheric data remain under- and interactive AI-driven analytics for
real-time monitoring. explored. enhanced scientific collaboration.
Operational Meteorol- AI-driven models refine weather fore- Lack of interpretability in AI forecasts, Hybrid AI-NWP models, physics-
ogy casts using ML and DL, integrating generalization issues under novel condi- constrained AI, and edge computing
multi-source data for extreme weather tions, and high computational costs. for real-time weather analysis and
event prediction and downscaling. enhanced extreme event forecasting.
Satellite Remote Sens- AI extracts features from vast satellite Processing large datasets, limited la- Explainable AI, hybrid AI-remote sens-
ing datasets, enhancing image classification, beled training data, and AI model trans- ing approaches, and improved annotated
hyperspectral analysis, and multi-modal parency for physical interpretation. datasets for enhanced AI training.
fusion for environmental monitoring.
Earth System DL improves climate simulations, re- Interpretability of DL models, maintain- Physics-informed neural networks
Modeling and Climate places conventional parameterizations, ing physical consistency, handling un- (PINNs), scalable DL models on
Change and enhances subgrid process modeling. certainty, and computational costs. HPC/cloud, and probabilistic AI
Unsupervised learning refines climate methods for uncertainty quantification.
variability analysis.

how land features like vegetation cover, topography, and soil state-of-the-art, challenges, and possible solutions for AI in
moisture feedback into weather systems and carbon exchange atmospheric sciences. While possible solutions are limited for
processes [140]. These integrative approaches support the de- short future solutions. In the next section, this paper will
velopment of more robust climate predictions and Earth system discuss wider and futuristic aspects of AI-driven solutions for
models by incorporating interactions that were traditionally atmospheric sciences in terms of a roadmap. The proposed
studied in isolation. roadmap serves as generic futuristic long-term solutions which
AI is also revolutionizing extreme weather event prediction will shape better future for atmospheric sciences and their
by improving the accuracy and timeliness of forecasts for hur- respective real-world applications.
ricanes, tornadoes, and heatwaves [150]. By processing large-
scale datasets from satellites and ground stations, AI-driven IV. ROADMAP OF AI-D RIVEN SOLUTIONS FOR
models enhance disaster preparedness and response efforts, ATMOSPHERIC S CIENCES
enabling better risk assessment and mitigation strategies [151].
The hydrological cycle, which is deeply connected to at- The integration of AI into atmospheric and environmental
mospheric processes, also benefits from AI applications. Ma- sciences presents unprecedented opportunities to enhance re-
chine learning techniques are being utilized to predict rainfall search, monitoring, and predictive capabilities. As the demand
patterns, river flows, and drought conditions, enabling more for real-time, high-accuracy data continues to grow, future
effective water resource management. Such advancements are research must concentrate on key directions that not only lever-
crucial for agriculture, urban planning, and ecosystem sus- age AI-driven solutions but also contribute to the advancement
tainability [152], [153]. In agricultural meteorology, AI-driven of AI itself.
models are used to predict weather patterns and assess their The complexity of atmospheric phenomena offers rich
impact on agricultural production. These models help optimize datasets that can significantly enhance AI research. Simultane-
crop yield forecasting and food security planning by analyzing ously, improving the quality and scale of these data is vital for
shifts in temperature, precipitation, and other meteorological a better understanding of atmospheric processes. One essential
factors that influence agriculture [154]. avenue for exploration is optimizing sensing infrastructure for
Computational atmospheric chemistry is another area where future atmospheric systems. Advances in signal processing,
AI plays a transformative role. ML and DL algorithms are anomaly detection, and self-calibrating sensors can improve
increasingly used to simulate complex chemical interactions in operational efficiency of sensing infrastructure while produc-
the atmosphere, predict pollutant transformations, and enhance ing high-quality data that serves as valuable training sets for
air quality models. These AI-driven advancements provide machine learning models. These innovations enable large-
valuable insights into the interplay between atmospheric chem- scale deployments, from low-cost sensors to sophisticated
istry, climate change, and air pollution control strategies [155]. instrumentation at research stations, fostering a more nuanced
understanding of atmospheric dynamics while also providing
a wealth of data to challenge AI algorithms.
H. Summary As the volume and complexity of atmospheric data in-
The discussions in the preceding subsections are synthesized crease, next-generation computing platforms for atmospheric
in Table II, which provides a concise overview of current data processing will play a crucial role. By combining high-
 13

performance computing, cloud computing, and edge-fog plat- frastructure to address the challenges of real-time, scalable,
forms, we can process massive datasets with minimal latency, and sustainable data collection.
which is vital for both atmospheric research and AI devel- 1) Self-sustaining monitoring networks: Energy inadequacy
opment. This capability allows researchers to conduct real- is a significant barrier to the long-term operation and deploy-
time simulations and analyses, pushing the boundaries of AI ment of atmospheric sensing systems, particularly in remote,
in processing and interpreting complex systems, thus driving inaccessible, or harsh environments. To address this challenge,
innovation in AI methodologies. At the same time, faster and there has been increasing interest in self-powered sensors capa-
more reactive data processing enhances atmospheric fields, ble of harvesting energy from their surroundings [156]. These
supporting applications from real-time weather forecasting to sensors can scavenge energy from environmental sources, such
disaster response and long-term climate modeling. as sunlight, radio frequency (RF) electromagnetic waves, ther-
Emerging directions in AI, referred to as advanced AI mal gradients, and mechanical energy, enabling sustainable,
methodologies, such as foundation models and explainable maintenance-free operations over extended periods [157].
AI, can also benefit from atmospheric sciences. Large-scale While solar and RF harvesting have long been investigated
foundation models trained on multi-modal datasets derived in the context of environmental monitoring [158], significant
from atmospheric research can uncover intricate patterns and challenges remain in transforming these technologies into self-
relationships that individual modalities may miss, improving sustained monitoring networks. Specifically, the quantities of
our understanding of atmospheric phenomena. In parallel, energy harvested tend to be relatively low, often sufficient only
the diverse and dynamic nature of atmospheric data presents for simple operations such as gathering individual readings
unique challenges that can inspire the development of new AI and transmitting them, rather than for storing and processing
algorithms and techniques. Moreover, incorporating explain- data. Additionally, sampling certain atmospheric compounds
able AI approaches ensures transparency and trust, particularly requires energy-intensive techniques that exceed the capabili-
in high-stakes applications such as climate predictions and ties of existing harvesting methods [85].
public health policy. Beyond solar and RF harvesting, recent advancements in
To effectively leverage these advancements, it is crucial to ocean wave energy conversion present promising possibilities
also support the end-users of the information through an AI- for powering ocean-based IoT devices. These are crucial for
driven transition, referred to as from data to action: analyt- continuous monitoring of marine environments and under-
ics, visualization, and automated feedback for environmental standing atmospheric-ocean interactions that significantly in-
decision-making. Atmospheric sciences benefit from AI-driven fluence climate systems and weather predictions [159]. By pro-
insights that translate into actionable outputs for decision- viding high-resolution, real-time data on oceanic conditions,
makers, scientists, and the public. Conversely, AI research these self-powered sensors enhance the quality of atmospheric
gains from the practical scenarios and case studies provided by models and improve forecasting accuracy, supplying valuable
atmospheric sciences, illustrating the real-world implications data for machine learning algorithms in AI applications.
of AI insights. For instance, automated feedback loops can Despite these innovations, substantial challenges persist.
enhance operational efficiency in optimizing indoor air quality Optimizing energy efficiency in sensors and processing sys-
or guiding policy interventions to mitigate harmful emissions, tems is critical to complementing energy harvesting, while
while also refining AI models based on user interactions and issues related to sensor size, cost, robustness, long-term sta-
outcomes. bility, and seamless integration with AI-driven data processing
Finally, the increase in data collection raises significant systems also need to be addressed [156]. Advancing self-
legislative concerns about data management, usage, and col- powered sensor technology requires a multidisciplinary ap-
lection practices. Data governance, privacy, and collaboration proach to improve the durability, adaptability, and energy
highlight the importance of ethical AI deployment and ro- efficiency of these devices. Continuous innovation in this field
bust governance frameworks, which are particularly relevant is vital for developing resilient and reliable sensing networks
in atmospheric sciences where data transparency is crucial. capable of meeting the demands of next-generation atmo-
By ensuring data privacy, compliance, and equitable access spheric monitoring systems. By integrating AI methodologies
to resources, interdisciplinary collaboration among scientists, and advanced data analytics, researchers can gain deeper
policymakers, and industry leaders can be fostered, driving insights into atmospheric phenomena, ultimately leading to
mutual benefits for both fields. more effective environmental management and policy-making.
Figure 7 summarizes these interconnected trends, illus- 2) Biodegradable environmental sensors: As sensor net-
trating how atmospheric sciences not only benefit from AI works continue to grow, the issue of electronic waste is
integration but also provide fertile ground for the advancement becoming increasingly critical and necessitating sustainable
of AI research itself. This reciprocal relationship holds the solutions for the disposal of electronic devices [160], [161].
potential to fuel the next generation of innovations in both This issue is particularly critical for atmospheric and envi-
atmospheric and AI sciences. ronmental monitoring as deployments scale up. One promis-
ing approach is the development of biodegradable sensors
made from environmentally friendly materials, which offer
A. Sensing and Communication Infrastructure a sustainable monitoring solution, especially in ecosystems
The future of atmospheric monitoring demands ground- vulnerable to pollution. These sensors are designed for short-
breaking innovations in the sensing and communication in- term deployments and naturally decompose after their opera-
 14

Fig. 7: Roadmap of AI-driven solutions and research potentials for atmospheric sciences

tional lifespan, significantly reducing their environmental im- actions, which is particularly relevant for domains such as
pact. Recent advancements have demonstrated the potential of disaster prevention or mitigation [164], [165]. The embedded
biodegradable and renewable materials, such as antennas made algorithms on edge sensors are adept at handling complex
from sustainable substrates for environmental sensing appli- tasks, including managing noisy data through real-time signal
cations [162]. For instance, chemical-based wearable sensors processing, implementing advanced anomaly detection sys-
have shown promise in detecting environmental pollutants, tems, and utilizing machine learning classifiers to address
but challenges remain in advancing these systems into next- and mitigate sensor drift. Self-calibration techniques further
generation environmental sensors that can be widely deployed. enhance measurement accuracy and extend sensor operational
From a materials perspective, a deeper understanding of the lifespans by automatically correcting sensor drift [166].
properties of nanomaterials is crucial for achieving biocompat- From a technical standpoint, edge devices face constraints
ibility, self-healing capabilities, and high performance at low such as limited computational and storage resources, particu-
costs. In addition, hybrid energy devices with low form factors larly when deploying DL models for real-time data stream pro-
are desirable for powering these sensors, as they should utilize cessing with low latency. Environmental challenges, including
multiple types of energy storage and harness energy from minimizing pollution hazards, traffic waste, resource consump-
diverse fuel sources to enable continuous operation. Ensuring tion, and energy usage, add further complexity. Furthermore,
high-speed, low-power real-time data communication is also cost-related concerns such as high initial capital investments,
vital, requiring the integration of advanced wireless tech- ongoing operational expenses, and the need to develop vi-
nologies, such as millimeter-wave frequency-based or optical able business models for integrating these technologies into
networks. Furthermore, biosynthesis-based green approaches atmospheric monitoring systems must also be addressed [167].
can be employed for the on-site remediation of environmental Finally, ensuring interoperability among different devices and
contaminants introduced by nanomaterials [163]. Addressing systems is vital to maximize the effectiveness of edge deploy-
these challenges and advancing biodegradable sensor tech- ments.
4) Innovative Satellite Technologies: CubeSats, a class of
nologies paves the way for greener atmospheric monitoring
nanosatellites, have emerged as a cost-effective alternative
systems that reduce the environmental footprint of sensor net-
to conventional satellites. These small, modular satellites are
works and enhance their compatibility with fragile ecosystems,
constructed using commercial off-the-shelf components, mak-
supporting sustainable environmental management.
ing them both affordable and adaptable for a wide range
3) AI-powered edge sensors: Integrating lightweight AI of applications [168]. Their scalability and flexibility have
models into edge devices enables real-time data processing positioned CubeSats as a transformative tool for atmospheric
directly at the source. These sensors can execute critical monitoring, enabling researchers to tackle challenges that
functions such as signal processing, anomaly detection, and traditional satellite systems often struggle to meet.
calibration locally. This reduces the need for extensive data One key advantage of CubeSats is their ability to commu-
transmission to centralized systems and facilitates immediate nicate effectively with one another and with ground stations,
 15

facilitating coordinated operations essential for many fields of more precise predictions and more effective disaster mitigation
atmospheric sciences. For example, tasks such as weather pre- strategies [176], [177].
diction, climate change monitoring, and disaster management The emergence of Quantum Computing as a Service
benefit from coordinated sampling [169]. The deployment of (QCaaS) within cloud platforms represents a transformative
CubeSat constellations offers a scalable solution for high- solution for atmospheric sciences, providing on-demand access
resolution atmospheric observations, effectively filling gaps to quantum resources without necessitating dedicated infras-
left by traditional systems. These constellations provide near tructure [178]. This hybrid approach facilitates the seamless
real-time insights into localized weather events [22] and urban integration of classical HPC approaches with quantum sys-
air quality information [170], thereby enhancing decision- tems, enabling the efficient execution of complex and compu-
making capabilities in environmental management. tationally intensive atmospheric algorithms. By synergizing the
Despite their potential, CubeSats face significant techni- strengths of both classical and quantum computing, researchers
cal and hardware-related challenges. Persistent issues include can unlock new capabilities for data analysis and predictive
power control, miniaturization, and configuration [171]. Their modeling, advancing our understanding and management of
small size limits the integration of advanced capabilities, atmospheric phenomena. Moreover, the multi-modal nature of
such as next-generation wireless communication systems and atmospheric datasets and the intricate patterns they encompass
sophisticated AI processing capabilities [169]. In addition, the can stimulate research in quantum-enabled AI and machine
compact design may restrict payload capacity, hindering the learning algorithms, opening avenues for developing innova-
deployment of advanced sensors necessary for comprehensive tive quantum algorithms.
atmospheric analysis. 2) Neuromorphic Computing: Modern computing systems
Addressing these challenges is crucial for enhancing the per- consume vast amounts of energy, rendering them unsustainable
formance and reliability of CubeSats, improving the quality of for the complex and data-intensive tasks necessary for atmo-
data they deliver, and ensuring they effectively support atmo- spheric modeling and environmental monitoring. While not
spheric sciences and environmental monitoring. Innovations in all computationally demanding processes require AI or DL,
power management systems, miniaturization techniques, and the widespread adoption of DL exacerbates energy concerns.
data processing methodologies can enable CubeSats to provide For instance, training a single DL model can consume up to
higher-resolution, real-time atmospheric data, thereby fueling 656.347 kilowatt-hours of energy, resulting in approximately
demand for advancements in AI processing techniques. By 626.155 pounds of CO2 emissions, equivalent to the total
overcoming these obstacles, CubeSats can solidify their role as lifetime carbon footprint of five automobiles [179]. Neu-
indispensable tools in environmental monitoring and enhanced romorphic computing has emerged as a promising solution
decision-making frameworks. to improve the sustainability of AI by developing systems
inspired by the structural and functional characteristics of the
human brain. This field aims to overcome the limitations of
B. Computing Infrastructure
conventional computing by replicating the parallelism, fault
Advancing atmospheric sciences requires robust and scal- tolerance, and energy efficiency observed in biological neu-
able computing infrastructure capable of handling vast datasets ral networks. Neuromorphic systems integrate hardware and
and complex simulations. As models grow in complexity and software architectures that mimic the behavior of neurons and
the demand for real-time analysis and forecasting increases, synapses, enabling more efficient processing of large datasets
innovations in computing ranging from high-performance with significantly lower energy consumption [180].
computing (HPC) to emerging paradigms like neuromorphic While neuromorphic computing has demonstrated improve-
and quantum computing will be essential to drive efficiency, ments in areas such as event-based data processing, adaptive
reduce energy consumption, and unlock new capabilities for control, constrained optimization, sparse feature regression,
environmental monitoring and predictive modeling. and graph search [181], its applications in atmospheric sci-
1) Quantum Computing: Quantum computing harnesses ences remain largely unexplored. Neuromorphic computing
the principles of superposition and entanglement to perform holds significant promise for advancing AI-driven atmospheric
computations at speeds unattainable by classical comput- modeling and monitoring systems by facilitating the train-
ers [172]. This capability enables quantum computers to ing of large-scale AI models with reduced energy demands.
address problems currently deemed intractable, even for the Furthermore, the integration of neuromorphic computing has
fastest classical systems, offering significant potential for ad- significant promise to improve the sustainability of complex
vancements in atmospheric sciences. The unique architecture AI models. For example, foundation models such as Mi-
of quantum computing allows for specialized algorithms that crosoft’s Aurora [182] and Google’s GenCast have demon-
can enhance climate modeling and weather prediction [173]. strated substantial benefits for weather forecasting and climate
Recent advancements have demonstrated the successful modeling [183]. Neuromorphic computing is also a promising
application of quantum computers in climate modeling and solution for low-power, smart edge processing in extreme
weather prediction [174]. For instance, numerical weather environments. This capability is particularly advantageous for
prediction (NWP) executed on quantum platforms has shown on-board AI processing in satellites, where energy efficiency
promise in enhancing computational speed and improving and real-time data analysis are critical [184]. By enabling AI-
forecast accuracy [175]. Furthermore, quantum machine learn- powered edge computing directly on satellites, neuromorphic
ing techniques can refine weather forecasting models, yielding computing reduces the necessity for continuous data transmis-
 16

sion to ground stations, thereby enhancing efficiency [185]. atmospheric research. However, atmospheric data also pose
3) DNA-based Molecular Computing and Storage: Signifi- challenges for FL. For example, atmospheric data are often
cant advances in biotechnology have enabled efficient manip- characterized by strong temporal correlations and are prone
ulation of deoxyribonucleic acid (DNA), making it possible to to distributional shifts [61], both of which complicate the
harness DNA as a programmable biological substrate. DNA use of conventional FL techniques. Another key issue is how
computing is a recently emerged field that builds on this to manage stragglers, nodes whose training contributions are
premise, leveraging DNA for computing, data storage, and delayed, without negatively impacting the final model. These
communication [186]. DNA operates with exceptional energy challenges are currently active areas of research [194], and
efficiency and possesses remarkable longevity and information access to atmospheric datasets can significantly contribute to
density, making it a highly attractive medium for a variety advancing this research.
of applications [186]. At the same time, the versatility of
DNA nanotechnology facilitates the design of programmable
C. Advanced AI Methodologies
DNA-based nanostructures and microstructures capable of
performing tasks ranging from molecular computing to data The next generation of AI in atmospheric sciences is ex-
storage [187]. pected to be driven by advancements in sophisticated models
Despite its significant potential, the application of DNA that can handle multimodal datasets and operate across dif-
computing in atmospheric sciences remains limited. One no- ferent spatial and temporal scales. This has the potential to
table exception is the use of spectral encoding and matching reveal phenomena that link different environmental systems
algorithms inspired by DNA computing to classify spectral and deepen our understanding of atmospheric processes.
signatures in hyperspectral remote sensing data [188]. This 1) Foundation Models and LLMS: Foundation models rep-
approach demonstrates the feasibility of DNA-based tech- resent a transformative leap in AI by enabling generalization
niques in addressing complex environmental data analysis across diverse tasks and domains through the utilization of
tasks. However, several challenges hinder the broader adop- vast training datasets. In atmospheric sciences, these models,
tion of DNA computing. Developing complex DNA-based trained on extensive repositories of climate, weather, and
devices that can function within live cells and regulate their remote sensing data, hold significant potential to enhance tasks
behavior is a significant obstacle. Moreover, as the number such as weather forecasting, environmental monitoring, and
of DNA strands required for computation grows exponentially anomaly detection [195]. For instance, developing universal
with problem complexity, improving the efficiency of DNA remote sensing foundation models that leverage millions of
synthesis remains a critical area of ongoing research [187]. spectral images can improve applications including multi-
Although DNA computing is still evolving, it holds immense label scene classification, semantic segmentation, and land-
promise for addressing key challenges in atmospheric sci- use change detection [196]. A notable example is the Aurora
ences, particularly in areas requiring large-scale data storage, foundation model, developed by Microsoft, which has been
energy-efficient processing, and cost-effective solutions. As trained on over a million hours of weather and climate data
advancements continue, DNA-based computing could become to produce high-fidelity operational forecasts across diverse
a transformative tool for managing the vast datasets associated atmospheric variables. Aurora has demonstrated exceptional
with atmospheric monitoring and modeling. performance, particularly in scenarios characterized by limited
4) Federated Learning with Distributed Computing Nodes: training data and extreme weather events.
Federated learning (FL) enables the training of AI models Large language models (LLMs) [60], a significant subset of
across decentralized data sources without transferring raw foundation models, are also gaining traction in atmospheric
data, preserving privacy and improving computational effi- and geoscience research. These models excel at interpreting
ciency [189]. FL can be highly valuable also in atmospheric complex textual data, generating coherent and contextually
sciences as it enables localized data processing without com- relevant responses, and facilitating interdisciplinary collabo-
promising security or data ownership. For instance, FL can ration. An example is the GeoGalactica model [197], which
be employed to analyze crowd-sensed weather data from extends conventional LLM capabilities by further pre-training
networks of low-cost, widely distributed weather stations, on geoscience texts and fine-tuning with domain-specific in-
facilitating precise local forecasts while safeguarding data struction datasets. This enables geoscientists to explore a wide
privacy [190]. FL has also shown promise in source term array of topics, generate hypotheses, and interpret complex
estimation, where deep neural networks identify unknown gas datasets more effectively. Due to their scale and sophistica-
leakage sources in urban environments [191]. Another example tion, foundation models possess the potential to accurately
is its use in remote sensing where FL can improve privacy, predict and forecast intricate scientific phenomena, such as
scalability, and classification accuracy [192]. new particle formation (NPF) and cyclone events, across
Beyond its privacy benefits, FL can significantly reduce varied geographical regions. By addressing both atmospheric
carbon emissions by decentralizing model training, leading processes and AI’s evolving challenges, these models help to
to lower overall energy consumption [193]. FL also holds drive forward advancements in both fields.
promise for generating global models that aggregate data from 2) Generative AI Models: Generative AI encompasses a
atmospheric monitoring stations, satellites, and edge sensors. class of algorithms and models designed to create new, pre-
Besides preserving data sovereignty, this approach enhances viously unseen data that closely resembles existing examples
scalability and lowers the environmental impact of AI-driven by learning the underlying patterns and structures present in
 17

the training ata [198]. This category includes various model climate modeling, weather forecasting, and disaster response,
types, such as generative adversarial networks (GANs), varia- the importance of explainability and interpretability will fur-
tional autoencoders (VAEs), and diffusion models [199]. These ther grow, as ensuring accountability and trust in these systems
models leverage advanced techniques to mimic statistical pro- is essential. Another important role for XAI is to help uncover
cesses, enabling the generation of synthetic data that augments relationships that could otherwise be unobservable. For exam-
datasets where labeled data are limited, thereby aiding in the ple, a recent study involved the use of XAI to uncover the
training of more robust AI models [198]. role of air pollutants, specifically O3 and NO2, as important
In atmospheric science, the GenCast model developed by contributors to Alzheimer’s disease mortality rates [203].
Google DeepMind [183] operates as a conditional diffusion Interactive visualization tools powered by XAI can further
model, capable of producing new samples from a given data enhance its benefits by enabling scientists to explore AI out-
distribution. GenCast has demonstrated its ability to generate puts, trace prediction pathways, and diagnose errors, thereby
high-resolution, probabilistic weather forecasts by learning refining atmospheric models. By exposing the inner workings
from historical climate data. By producing ensemble sim- of AI systems, XAI not only enhances the credibility of
ulations that capture uncertainty and rare atmospheric phe- forecasts but also facilitates collaborative research by allowing
nomena, GenCast surpasses conventional numerical weather domain experts to question and refine model behavior.
prediction (NWP) methods, enhancing disaster prediction and 4) Physics-Informed Machine Learning (PIML): Physics-
mitigation capabilities. More generally, while diffusion models informed machine learning (PIML) integrates domain knowl-
have primarily focused on web content such as text, images, edge and physical principles directly into machine learning
and video, they hold significant potential for bridging multi- models, enhancing their accuracy and interpretability [204].
modal datasets used in atmospheric sciences. This approach, also referred to as knowledge-guided machine
A key advantage of generative AI is its effectiveness in learning (KGML), ensures that underlying physical constraints
addressing the challenge of limited labeled data, making it par- are respected during the learning process, leading to more
ticularly adept at supporting unsupervised and semi-supervised robust and reliable predictions [205]. A prominent example
learning. By generating synthetic data that replicates real- of PIML is the development of Physics-Informed Neural
world conditions, these models improve performance when Networks (PINNs), a class of machine learning techniques
labeled datasets are scarce. This characteristic is especially designed to solve problems governed by partial differential
valuable in semi-supervised learning frameworks, where small equations [206]. PINNs have been successfully applied to em-
amounts of labeled data are supplemented with large volumes ulate, downscale, and forecast weather and climate processes,
of unlabeled data. Notable applications of generative AI in demonstrating their potential to enhance both computational
atmospheric sciences include extreme weather prediction [58] efficiency and predictive performance [207].
and hyperspectral image classification in remote sensing [126]. An emerging innovation is the integration of PINN ar-
These applications highlight the transformative potential of chitecture with foundation models to embed physical laws
generative models in advancing our understanding of complex into large-scale AI systems. This integration can reduce the
atmospheric phenomena and improving predictive capabilities. complexity of foundation models while maintaining accuracy
3) Explainable AI (XAI): Explainable Artificial Intelligence and generalizability. For instance, ClimaX, a foundation model
(XAI) encompasses a set of techniques and tools designed to developed by Microsoft for weather forecasting and climate
make machine learning models more transparent, interpretable, projection, features a large and complex architecture [208].
and trustworthy for human users [200]. This includes vari- By incorporating PINNs into ClimaX, the model’s size and
ous methods, such as model-agnostic approaches like LIME computational demands could be scaled down without compro-
(Local Interpretable Model-agnostic Explanations) and SHAP mising performance, offering a more efficient and interpretable
(SHapley Additive exPlanations), as well as inherently inter- approach for atmospheric and climate science. The synergy
pretable models like decision trees and linear regression. In between PIML and traditional ML can help improve the
atmospheric sciences, XAI plays a critical role in demystifying performance of ML models while at the same time address
complex AI models, enabling researchers to trace model the challenges posed by the complex and chaotic nature
predictions back to the influencing factors and increasing trust of atmospheric phenomena. By leveraging the strengths of
in AI-driven insights. For instance, XAI has been employed both domains, researchers can develop more accurate and
to predict cirrus cloud formation, revealing the relationships interpretable models that simultaneously reduce computational
between meteorological and aerosol conditions that affect costs. Conversely, the atmospheric sciences provide a rich
cirrus properties [201]. This application illustrates how XAI domain governed by clear physical laws, presenting significant
can enhance our understanding of atmospheric processes and opportunities to foster the future development of physics-
improve model accuracy. informed machine learning models.
XAI techniques provide insights into how models generate 5) Neurosymbolic AI for Causal Inference: Neurosymbolic
predictions, offering transparency for scientists and policymak- AI, which combines neural networks with symbolic reasoning,
ers [98]. This transparency is crucial for fostering confidence represents a significant advancement in causal discovery and
in AI-driven climate models, particularly in regulatory and interpretation. By embedding symbolic rules into neural mod-
policy-driven environments where interpretability is essential els, neurosymbolic AI can infer cause-and-effect relationships
for informed decision-making [202]. AI becomes increasingly from observational data, thereby enhancing the reliability of
integrated into high-stakes atmospheric applications, such as AI systems used for climate modeling and environmental
 18

prediction [99]. This approach is particularly promising for personalized climate knowledge to empower informed action
detecting complex feedback loops in atmospheric systems, across society [215], [216].
such as the interplay between aerosols and cloud formation. Beyond global-scale assessments, digital twins can aid en-
ergy providers in designing deep mantle geothermal energy
projects and assist the reinsurance industry in testing risk
D. From Data to Action
mitigation strategies and disaster management scenarios [217].
The future of atmospheric sciences hinges on ability to Their applications extend to agriculture, where they monitor
transform complex environmental data into actionable insights. crop health and optimize resource use. At the urban scale,
By utilizing advanced AI-driven analytics, visualization tech- digital twins can be downscaled to enhance city systems by
niques, and automated feedback systems, decision-makers can integrating advanced visualization tools, such as augmented
respond more effectively to environmental challenges. These reality (AR) and holographic displays, optimizing air qual-
innovations are crucial for identifying pressing issues and ity management, and supporting smart city initiatives [218].
developing more effective, actionable policies that promote Despite their transformative potential, the implementation of
sustainability and resilience. digital twins faces challenges such as the need for high-quality
1) Real-Time Analytics and Automated Feedback Loops: data, data privacy concerns, and the complexity of accurately
Real-time analytics platforms facilitate immediate processing modeling dynamic systems. Addressing these challenges is
of environmental data, and provide decision-makers with ac- essential for realizing the full potential of digital twins in
tionable insights. Adaptive systems can dynamically adjust tackling environmental and urban challenges.
sensor calibration processes and modify the data resolution 3) Citizen Science and Explainable AI for Public Engage-
based on current conditions. This can improve monitoring ment: Citizen science initiatives harness crowdsourced data
precision. AI-driven feedback loops help bridge the gap be- from citizens, their mobile devices and diverse IoT-enabled
tween data collection and action by autonomously controlling sensor devices to augment traditional monitoring systems
environmental actuators. For example, indoor air quality mon- while fostering public awareness and participation. These
itoring can adjust Heating, Ventilation, and Air Conditioning initiatives enable everyday citizens to contribute to scientific
(HVAC) operations to maintain optimal conditions [209]. research and environmental monitoring efforts. AI models,
Similarly, flood barriers can be automatically deployed in including ML algorithms, validate and process this data,
response to real-time weather data [210], while urban traffic ensuring reliability for decision-making. Explainable AI is also
systems can optimize traffic flows to reduce emissions [211]. necessary to enhance the transparency regarding how models
AI-driven analytics also enhance the management of re- generate predictions, helping policymakers and stakeholders
newable energy systems. Hyper-local weather estimation and understand AI-driven insights and fostering trust in these
forecasting improve the efficiency of solar and wind energy systems [98], [202]. Gamified platforms are another notable
systems by optimizing energy distribution and storage [30]. method for improving public awareness. These engage the
Moreover, Reinforcement Learning (RL) has demonstrated public by simulating environmental scenarios and encourag-
great potential in optimizing various systems by making ing sustainable behavior [219]. By incorporating elements
sequential decisions in dynamic environments. For example, of play and competition, these platforms make participation
RL has been successfully employed to optimize energy con- more appealing and effective. Furthermore, citizen science can
sumption in HVAC systems, adjusting operations based on be amplified through educational workshops and community
weather predictions and building occupancy patterns [212]. In events that emphasize collective action for environmental
addition, RL frameworks can dynamically allocate resources sustainability.
during extreme weather events, such as redirecting power to
critical infrastructure during storms or floods [213].
These applications are examples of how real-time analyt- E. Ethical, Collaborative, and Regulatory Considerations
ics and automated feedback systems can evolve from data
interpretation into proactive frameworks. By enabling timely The integration of AI into atmospheric sciences presents not
interventions, these technologies drive improvements in sus- only technological advancements but also critical challenges
tainability and enhance environmental resilience. in ethics, collaboration, and regulation [220]. Addressing these
2) Digital Twin of Earth and Cognitive Systems: Digital aspects is essential to ensure responsible and impactful use of
twins of the Earth are sophisticated digital replicas that en- AI in monitoring and predictive systems.
compass various scales and domains of the Earth system, 1) Data Governance and Privacy: Atmospheric monitoring
enabling comprehensive monitoring, forecasting, and assess- involves large-scale data collection from diverse public and
ment of environmental processes and the consequences of private sources, raising concerns about the ethical handling
human interventions [214]. These systems empower policy- of sensitive information. Ensuring data governance and pri-
makers and scientists to explore complex phenomena, such as vacy requires the adoption of policies that prioritize data
storm evolution, and explore ”what-if” scenarios for climate anonymization, secure storage, and transparent usage [221].
interventions, urban planning, and disaster response. They can These measures safeguard individual and organizational pri-
facilitate data-driven decisions regarding emission control and vacy while maintaining data integrity [15]. Addressing these
resource management, supporting initiatives like the Earth concerns is critical to building trust in AI-driven systems and
Virtualization Engines (EVE) program, which aims to provide facilitating their widespread adoption in atmospheric sciences.
 19

Fig. 8: Emerging trends and future research directions in AI for atmospheric science and its related infrastructure

2) Interdisciplinary Collaboration: The interdisciplinary compliance with environmental regulations. AI technologies

nature of AI and atmospheric science integration necessitates can also play a proactive role in policy-making by offering
collaboration between atmospheric scientists, AI experts, pol- real-time insights and actionable recommendations, allowing
icymakers, and industry stakeholders [222]. Joint efforts are policymakers to respond swiftly to environmental challenges
essential to tackle complex environmental challenges, ensuring while maintaining regulatory alignment [225].
that AI tools are designed to meet the needs of all stakeholders.
Partnerships with technology providers, regulatory agencies,
F. Summary
and international organizations can further drive AI adoption
on a global scale, creating solutions that effectively address en- As our overview has consistently highlighted, the integration
vironmental monitoring and climate change mitigation [223]. of advanced AI techniques into atmospheric sciences holds sig-
3) Regulatory Frameworks and Policy Support: Advance- nificant promise for enhancing our understanding of complex
ments in AI require robust regulatory frameworks to ensure environmental phenomena. The proposed roadmap, illustrated
transparency, accountability, and ethical usage in environmen- in Figure 8, encompasses the entire computing continuum,
tal monitoring systems [224]. Governments and international enhancing the analysis of atmospheric datasets and providing
organizations must establish policies that support responsible benefits throughout the entire data life cycle from collection
AI development, enable cross-border data sharing, and ensure to end-user applications.
 20

In terms of deployment, the AI roadmap envisions integra- the same time, we have highlighted opportunities on how
tion across a broad computing spectrum, from low-end sensor atmospheric sciences can boost AI research by presenting
devices to fog and edge nodes, as well as cloud and high- unique challenges that push the boundaries of current methods.
performance computing (HPC) platforms. Data-producing de- By leveraging measurements from diverse sources, including
vices can leverage federated learning to ensure data privacy satellite remote sensing, ground-based measurements, and IoT
and sovereignty while utilizing calibration models to improve devices, AI technologies are transformed our understanding of
measurement quality and derive localized insights. As emerg- complex environmental systems and revolutionizing air quality
ing paradigms such as DNA computing and neuromorphic monitoring, operational meteorology, and climate modeling.
computing mature, they present opportunities to further en- We identified key applications of AI across key atmospheric
hance computational efficiency. HPC and large-scale cloud science domains, including air quality monitoring and model-
platforms facilitate the analysis of vast datasets using foun- ing, in-situ atmospheric observations, operational meteorology,
dation models and generative AI, which improve the accuracy satellite remote sensing, Earth system modeling, and climate
of weather forecasts and climate models. These AI techniques change. Our analysis highlighted the current state of the field,
excel at handling datasets comprising of multiple modalities ongoing challenges, and potential solutions. While AI offers
and limited labels, common characteristics of measurements in unprecedented opportunities to enhance predictive capabilities,
many atmospheric domains. Looking further ahead, quantum improve climate model accuracy, and enable more efficient
computing has the potential to revolutionize the processing of environmental monitoring, data scarcity, model interpretability,
atmospheric data, enabling the discovery of complex patterns development of hybrid AI-physics models and other challenges
that are currently beyond our reach. However, the successful must be first addressed. Looking ahead, we also presented a
adoption of these AI techniques hinges on building trust research roadmap that outlines key roles emerging technolo-
in AI-derived insights, necessitating the implementation of gies, including foundation models and quantum computing,
explainable AI to trace decision-making processes and the use can play for future innovations to enable the next generation
of physical computing techniques to ensure they adhere with of research advances in atmospheric sciences.
physical laws governing atmospheric processes. To conclude, the integration of AI into atmospheric sciences
Conversely, advancements in atmospheric sciences present not only promises to improve our understanding and manage-
unique challenges and opportunities for AI research, especially ment of environmental phenomena but also offers a pathway
concerning the processing of multi-modal data and ensuring toward more sustainable interactions with our planet. By
the robustness of AI models against variations in data char- fostering collaboration between AI and atmospheric research,
acteristics. The complexity of atmospheric data sources, in- we can drive forward advancements that will benefit both
cluding satellite imagery, sensor readings, and climate models, fields and contribute to effective climate change mitigation
necessitates AI systems to be capable of effectively integrating and adaptation strategies.
and processing heterogeneous data types. This demand pushes
the boundaries of AI research, challenging current methods. As ACKNOWLEDGMENT
researchers work to enhance the resilience of AI models in the
face of unpredictable atmospheric phenomena, they are likely This project is supported by the Research Council of Finland
to uncover new approaches that improve AI’s generalizability through the Academy Research Fellowship projects (Grant
and performance, enriching the field of AI itself. No. 355330 and 362594), an Academy Project (Grant No.
The implications of integrating AI into atmospheric sci- 339614), and the University Profiling funding initiative In-
ences extend well beyond academic advancements. Enhanced terEarth (Grant No. 353218). Additional financial support from
weather forecasting and climate modeling can lead to better the Research Council of Finland is gratefully acknowledged
preparedness for extreme weather events, helping to save lives via the Atmosphere and Climate Competence Center (Grant
and reduce economic losses. AI-driven environmental moni- No. 337549, 357902, and 359340), as well as through the Re-
toring systems can promote sustainable practices in sectors search Infrastructure of the Institute for Atmospheric and Earth
such as agriculture, energy, and urban planning, increasing System Research – INAR RI (Grant No. 367739). We also
societal resilience. Citizen science initiatives that utilize these acknowledge support from the RI-URBANS project (Research
technologies enhance public engagement while empowering Infrastructures Services Reinforcing Air Quality Monitoring
communities to take informed actions regarding environmental Capacities in European Urban and Industrial Areas), funded by
issues. As we confront the complexities of climate change, the European Union’s Horizon 2020 research and innovation
the responsible integration of AI into atmospheric sciences programme under the Green Deal initiative (Grant Agreement
can facilitate informed policy-making and collaborative efforts No. 101036245).
aimed at achieving environmental sustainability, helping to We acknowledge the use of OpenAI’s ChatGPT-4
shape a more resilient future for both society and the planet. (https://chat.openai.com/) to refine the academic language and
coherence of selected sections.

V. C ONCLUSION
R EFERENCES
We examined the transformative potential AI can have for
[1] B. Fu, Y. Liu, Y. Li, C. Wang, C. Li, W. Jiang, T. Hua, and W. Zhao,
atmospheric sciences, emphasizing its potential to enhance “The research priorities of resources and environmental sciences,”
our understanding of complex environmental systems. At Geography and Sustainability, vol. 2, no. 2, pp. 87–94, 2021.
 21

[2] K. Abbass, M. Z. Qasim, H. Song, M. Murshed, H. Mahmood, and Sciences, T. C. Vance, N. Merati, C. Yang, and M. Yuan, Eds.
I. Younis, “A review of the global climate change impacts, adapta- Academic Press, 2016, pp. 137–161.
tion, and sustainable mitigation measures,” Environmental Science and [22] H. B. Bluestein, F. H. Carr, and S. J. Goodman, “Atmospheric obser-
Pollution Research, vol. 29, no. 28, pp. 42 539–42 559, 2022. vations of weather and climate,” Atmosphere-Ocean, vol. 60, no. 3-4,
[3] World Meteorological Organization (WMO). (2024) Devastating pp. 149–187, 2022.
rainfall hits spain: yet another flood-related disaster. Available: [23] P. Hari, E. Nikinmaa, T. Pohja, E. Siivola, J. Bäck, T. Vesala, and
https://wmo.int/media/news/devastating-rainfall-hits-spain-yet- M. Kulmala, “Station for measuring ecosystem-atmosphere relations:
another-flood-related-disaster, Accessed: 2025-01-09. Smear,” Physical and Physiological Forest Ecology, pp. 471–487, 2013.
[4] Copernicus Climate Change Service (C3S). (2024) Copernicus: Second [24] M. A. Zaidan, N. H. Motlagh, P. L. Fung, D. Lu, H. Timonen, J. Kuula,
warmest november globally confirms expectation of 2024 as warmest J. V. Niemi, S. Tarkoma, T. Petäjä, M. Kulmala, and T. Hussein,
year. Available: https://climate.copernicus.eu/copernicus-second- “Intelligent calibration and virtual sensing for integrated low-cost air
warmest-november-globally-confirms-expectation-2024-warmest- quality sensors,” IEEE Sensors Journal, vol. 20, no. 22, pp. 13 638–
year, Accessed: 2025-01-09. 13 652, 2020.
[5] A. C. O’Regan and M. M. Nyhan, “Towards sustainable and net-zero [25] P. Laj, C. Lund Myhre, V. Riffault, V. Amiridis, H. Fuchs,
cities: A review of environmental modelling and monitoring tools for K. Eleftheriadis, T. Petäjä, T. Salameh, N. Kivekäs, E. Juurola,
optimizing emissions reduction strategies for improved air quality in G. Saponaro, S. Philippin, C. Cornacchia, L. Alados Arboledas,
urban areas,” Environmental Research, vol. 231, p. 116242, 2023. H. Baars, A. Claude, M. De Mazière, B. Dils, M. Dufresne,
[6] C. Vaduva, M. Iapaolo, and M. Datcu, “A scientific perspective on big N. Evangeliou, O. Favez, M. Fiebig, M. Haeffelin, H. Herrmann,
data in earth observation,” Principles of Data Science, pp. 155–188, K. Höhler, N. Illmann, A. Kreuter, E. Ludewig, E. Marinou, O. Möhler,
2020. L. Mona, L. Eder Murberg, D. Nicolae, A. Novelli, E. O’Connor,
[7] T. C. Vance, T. Huang, and K. A. Butler, “Big data in earth science: K. Ohneiser, R. M. Petracca Altieri, B. Picquet-Varrault, D. van
Emerging practice and promise,” Science, vol. 383, no. 6688, p. Pinxteren, B. Pospichal, J.-P. Putaud, S. Reimann, N. Siomos, I. Stach-
eadh9607, 2024. lewska, R. Tillmann, K. A. Voudouri, U. Wandinger, A. Wieden-
[8] X. Zhang, Y. Zhou, and J. Luo, “Deep learning for processing and sohler, A. Apituley, A. Comerón, M. Gysel-Beer, N. Mihalopoulos,
analysis of remote sensing big data: A technical review,” Big Earth N. Nikolova, A. Pietruczuk, S. Sauvage, J. Sciare, H. Skov, T. Svendby,
Data, vol. 6, no. 4, pp. 527–560, 2022. E. Swietlicki, D. Tonev, G. Vaughan, V. Zdimal, U. Baltensperger, J.-F.
[9] M. Talebkhah, A. Sali, M. Marjani, M. Gordan, S. J. Hashim, and F. Z. Doussin, M. Kulmala, G. Pappalardo, S. Sorvari Sundet, and M. Vana,
Rokhani, “Iot and big data applications in smart cities: recent advances, “Aerosol, Clouds and Trace Gases Research Infrastructure (ACTRIS):
challenges, and critical issues,” IEEE Access, vol. 9, pp. 55 465–55 484, The European research infrastructure supporting atmospheric science,”
2021. Bulletin of the American Meteorological Society, vol. 105, no. 7, p.
[10] Y. Hajjaji, W. Boulila, I. R. Farah, I. Romdhani, and A. Hussain, “Big E1098–E1136, Jul. 2024.
data and iot-based applications in smart environments: A systematic [26] J. Heiskanen, C. Brümmer, N. Buchmann, C. Calfapietra, H. Chen,
review,” Computer Science Review, vol. 39, p. 100318, 2021. B. Gielen, T. Gkritzalis, S. Hammer, S. Hartman, M. Herbst, I. A.
[11] A. Salam, “Internet of Things for environmental sustainability and Janssens, A. Jordan, E. Juurola, U. Karstens, V. Kasurinen, B. Kruijt,
climate change,” in Internet of Things for sustainable community devel- H. Lankreijer, I. Levin, M.-L. Linderson, D. Loustau, L. Merbold, C. L.
opment: Wireless Communications, Sensing, and Systems. Springer, Myhre, D. Papale, M. Pavelka, K. Pilegaard, M. Ramonet, C. Reb-
2024, pp. 33–69. mann, J. Rinne, L. Rivier, E. Saltikoff, R. Sanders, M. Steinbacher,
[12] S. Poupry, K. Medjaher, and C. Béler, “Data reliability and fault T. Steinhoff, A. Watson, A. T. Vermeulen, T. Vesala, G. Vı́tková, and
diagnostic for air quality monitoring station based on low cost sensors W. Kutsch, “The integrated carbon observation system in Europe,”
and active redundancy,” Measurement, vol. 223, p. 113800, 2023. Bulletin of the American Meteorological Society, vol. 103, no. 3, p.
[13] R. S. Sokhi, N. Moussiopoulos, A. Baklanov, J. Bartzis, I. Coll, E855–E872, Mar. 2022.
S. Finardi, R. Friedrich, C. Geels, T. Grönholm, T. Halenka et al., [27] B. Zhou, S. Zhang, R. Xue, J. Li, and S. Wang, “A review of
“Advances in air quality research–current and emerging challenges,” space-air-ground integrated remote sensing techniques for atmospheric
Atmospheric Chemistry and Physics Discussions, vol. 2021, pp. 1–133, monitoring,” Journal of Environmental Sciences, vol. 123, pp. 3–14,
2021. 2023.
[14] S. Zhang, Y. Xue, X. Zhou, X. Zhang, W. Liu, K. Li, and R. Liu, [28] Q. Zhao, L. Yu, Z. Du, D. Peng, P. Hao, Y. Zhang, and P. Gong, “An
“State of the art: high-performance and high-throughput computing overview of the applications of earth observation satellite data: impacts
for remote sensing big data,” IEEE Geoscience and Remote Sensing and future trends,” Remote Sensing, vol. 14, no. 8, p. 1863, 2022.
Magazine, vol. 10, no. 4, pp. 125–149, 2022. [29] N. Saeed, T. Y. Al-Naffouri, and M.-S. Alouini, “Around the world of
[15] D. Tuia, R. Roscher, J. D. Wegner, N. Jacobs, X. Zhu, and G. Camps- IoT/climate monitoring using Internet of X-Things,” IEEE Internet of
Valls, “Toward a collective agenda on AI for earth science data Things Magazine, vol. 3, no. 2, pp. 82–83, 2020.
analysis,” IEEE Geoscience and Remote Sensing Magazine, vol. 9, [30] M. A. Zaidan, N. H. Motlagh, B. Zakeri, T. Petäjä, M. Kulmala,
no. 2, pp. 88–104, 2021. and S. Tarkoma, “Irmaset: Intelligent weather forecaster system for
[16] M. A. Zaidan, Y. Xie, N. H. Motlagh, B. Wang, W. Nie, P. Nurmi, hyperlocal renewable energies,” IEEE Consumer Electronics Magazine,
S. Tarkoma, T. Petäjä, A. Ding, and M. Kulmala, “Dense air quality vol. 13, no. 5, pp. 61–74, 2024.
sensor networks: Validation, analysis, and benefits,” IEEE Sensors [31] Y. Hashmy, Z. U. Khan, F. Ilyas, R. Hafiz, U. Younis, and T. Tauqeer,
Journal, vol. 22, no. 23, pp. 23 507–23 520, 2022. “Modular air quality calibration and forecasting method for low-cost
[17] I. Ahmed, G. Jeon, and F. Piccialli, “From Artificial Intelligence to sensor nodes,” IEEE Sensors Journal, vol. 23, no. 4, pp. 4193–4203,
explainable Artificial intelligence in industry 4.0: a survey on what, 2023.
how, and where,” IEEE Transactions on Industrial Informatics, vol. 18, [32] N. H. Motlagh, M. A. Zaidan, R. Morabito, P. Nurmi, and S. Tarkoma,
no. 8, pp. 5031–5042, 2022. Towards Large-Scale IoT Deployments in Smart Cities: Requirements
[18] P. D. Dueben, M. G. Schultz, M. Chantry, D. J. Gagne, D. M. Hall, and Challenges. Cham: Springer Nature Switzerland, 2024, pp. 105–
and A. McGovern, “Challenges and benchmark datasets for machine 129.
learning in the atmospheric sciences: Definition, status, and outlook,” [33] S. Salcedo-Sanz, P. Ghamisi, M. Piles, M. Werner, L. Cuadra,
Artificial Intelligence for the Earth Systems, vol. 1, no. 3, p. e210002, A. Moreno-Martı́nez, E. Izquierdo-Verdiguier, J. Muñoz-Marı́,
2022. A. Mosavi, and G. Camps-Valls, “Machine learning information fusion
[19] A. Petzold, U. Bundke, A. Hienola, P. Laj, C. Lund Myhre, A. Ver- in earth observation: A comprehensive review of methods, applications
meulen, A. Adamaki, W. Kutsch, V. Thouret, D. Boulanger et al., and data sources,” Information Fusion, vol. 63, pp. 256–272, 2020.
“Opinion: New directions in atmospheric research offered by research [34] P. Yue, B. Shangguan, L. Hu, L. Jiang, C. Zhang, Z. Cao, and
infrastructures combined with open and data-intensive science,” Atmo- Y. Pan, “Towards a training data model for artificial intelligence in
spheric Chemistry and Physics, vol. 24, no. 9, pp. 5369–5388, 2024. earth observation,” International Journal of Geographical Information
[20] M. Kulmala, “Build a global earth observatory,” Nature, vol. 553, no. Science, vol. 36, no. 11, pp. 2113–2137, 2022.
7686, pp. 21–23, 2018. [35] S. S. Gill, H. Wu, P. Patros, C. Ottaviani, P. Arora, V. C. Pujol,
[21] W. Li, H. Shao, S. Wang, X. Zhou, and S. Wu, “Chapter 9 - A2CI: A D. Haunschild, A. K. Parlikad, O. Cetinkaya, H. Lutfiyya et al., “Mod-
cloud-based, service-oriented geospatial cyberinfrastructure to support ern computing: Vision and challenges,” Telematics and Informatics
atmospheric research,” in Cloud Computing in Ocean and Atmospheric Reports, p. 100116, 2024.
 22

[36] N. Wedi, P. Bauer, I. Sandu, J. Hoffmann, S. Sheridan, R. Cereceda, M. Suttie, F. Vamborg, J. Varndell, S. Vermoote, X. Yang, and J. G.
T. Quintino, D. Thiemert, and T. Geenen, “Destination Earth: High- de Marcilla, “The copernicus climate change service: climate science
performance computing for weather and climate,” Computing in Sci- in action,” Bulletin of the American Meteorological Society, vol. 103,
ence & Engineering, vol. 24, no. 6, pp. 29–37, 2022. no. 12, pp. E2669–E2687, 2022.
[37] J. Michalakes, HPC for Weather Forecasting. Springer International [56] N. Jones, “How machine learning could help to improve climate
Publishing, 2020, p. 297–323. forecasts,” Nature, vol. 548, no. 7668, 2017.
[38] M. Alvanos and T. Christoudias, “Accelerating atmospheric chemical [57] C. Wong, “How AI is improving climate forecasts,” Nature, vol. 628,
kinetics for climate simulations,” IEEE Transactions on Parallel and pp. 710–712, 2024.
Distributed Systems, vol. 30, no. 11, pp. 2396–2407, 2019. [58] E. Racah, C. Beckham, T. Maharaj, S. E. Kahou, Prabhat, and C. Pal,
[39] J. Kukkonen, T. Olsson, D. M. Schultz, A. Baklanov, T. Klein, A. I. “Extreme weather: a large-scale climate dataset for semi-supervised
Miranda, A. Monteiro, M. Hirtl, V. Tarvainen, M. Boy, V.-H. Peuch, detection, localization, and understanding of extreme weather events,”
A. Poupkou, I. Kioutsioukis, S. Finardi, M. Sofiev, R. Sokhi, K. E. J. in Proceedings of the 31st International Conference on Neural Informa-
Lehtinen, K. Karatzas, R. San José, M. Astitha, G. Kallos, M. Schaap, tion Processing Systems, ser. NIPS’17. Red Hook, NY, USA: Curran
E. Reimer, H. Jakobs, and K. Eben, “A review of operational, regional- Associates Inc., 2017, p. 3405–3416.
scale, chemical weather forecasting models in Europe,” Atmospheric [59] C. Bai, M. Zhang, J. Zhang, J. Zheng, and S. Chen, “Lscidmr: Large-
Chemistry and Physics, vol. 12, no. 1, pp. 1–87, 2012. scale satellite cloud image database for meteorological research,” IEEE
[40] W. Voorsluys, J. Broberg, and R. Buyya, “Introduction to cloud Transactions on Cybernetics, vol. 52, no. 11, pp. 12 538–12 550, 2021.
computing,” Cloud computing: Principles and Paradigms, pp. 1–41, [60] F. Sarhaddi, N. T. Nguyen, A. Zuniga, P. Hui, S. Tarkoma, H. Flores,
2011. and P. Nurmi, “Llms and iot: A comprehensive survey on large
[41] H. M. Monti, A. R. Butt, and S. S. Vazhkudai, “Catch: A cloud-based language models and the internet of things,” Authorea Preprints, 2025.
adaptive data transfer service for hpc,” in 2011 IEEE International [61] K. Aula, E. Lagerspetz, P. Nurmi, and S. Tarkoma, “Evaluation of low-
Parallel and Distributed Processing Symposium, 2011, pp. 1242–1253. cost air quality sensor calibration models,” ACM transactions on sensor
[42] D. Montes, J. A. Añel, D. C. Wallom, P. Uhe, P. V. Caderno, and T. F. networks, vol. 18, no. 4, pp. 1–32, 2022.
Pena, “Cloud computing for climate modelling: Evaluation, challenges [62] Q. Zhao, Y. Liu, W. Yao, and Y. Yao, “Hourly rainfall forecast
and benefits,” Computers, vol. 9, no. 2, p. 52, 2020. model using supervised learning algorithm,” IEEE Transactions on
[43] M. Amani, A. Ghorbanian, S. A. Ahmadi, M. Kakooei, A. Moghimi, Geoscience and Remote Sensing, vol. 60, pp. 1–9, 2021.
S. M. Mirmazloumi, S. H. A. Moghaddam, S. Mahdavi, M. Ghahre- [63] S. M. Saad, A. M. Andrew, A. Y. M. Shakaff, A. R. M. Saad, A. M. Y. .
manloo, S. Parsian, Q. Wu, and B. Brisco, “Google Earth Engine Kamarudin, and A. Zakaria, “Classifying sources influencing indoor air
cloud computing platform for remote sensing big data applications: quality (IAQ) using artificial neural network (ANN),” Sensors, vol. 15,
A comprehensive review,” IEEE Journal of Selected Topics in Applied no. 5, pp. 11 665–11 684, 2015.
Earth Observations and Remote Sensing, vol. 13, pp. 5326–5350, 2020. [64] P. Govender and V. Sivakumar, “Application of K-means and hier-
[44] C. Yang, Q. Huang, Z. Li, K. Liu, and F. Hu, “Big data and cloud archical clustering techniques for analysis of air pollution: A review
computing: innovation opportunities and challenges,” International (1980–2019),” Atmospheric Pollution Research, vol. 11, no. 1, pp. 40–
Journal of Digital Earth, vol. 10, no. 1, pp. 13–53, 2017. 56, 2020.
[45] N. A. Angel, D. Ravindran, P. D. R. Vincent, K. Srinivasan, and Y.-C.
[65] K. B. Shaban, A. Kadri, and E. Rezk, “Urban air pollution monitoring
Hu, “Recent advances in evolving computing paradigms: Cloud, edge,
system with forecasting models,” IEEE Sensors Journal, vol. 16, no. 8,
and fog technologies,” Sensors, vol. 22, no. 1, p. 196, 2021.
pp. 2598–2606, 2016.
[46] W. Wang, C. Feng, B. Zhang, and H. Gao, “Environmental monitoring
[66] P. L. Fung, S. Sillanpää, J. V. Niemi, A. Kousa, H. Timonen, M. A.
based on fog computing paradigm and internet of things,” IEEE Access,
Zaidan, E. Saukko, M. Kulmala, T. Petäjä, and T. Hussein, “Improving
vol. 7, pp. 127 154–127 165, 2019.
the current air quality index with new particulate indicators using a
[47] L. Kong, J. Tan, J. Huang, G. Chen, S. Wang, X. Jin, P. Zeng, M. Khan,
robust statistical approach,” Science of the Total Environment, vol. 844,
and S. K. Das, “Edge-computing-driven Internet of Things: A survey,”
p. 157099, 2022.
ACM Computing Surveys, vol. 55, no. 8, pp. 1–41, 2022.
[48] H. A. B. Salameh, M. F. Dhainat, and E. Benkhelifa, “An end-to-end [67] E. Von Schneidemesser, C. Driscoll, H. E. Rieder, and L. D. Schiferl,
early warning system based on wireless sensor network for gas leakage “How will air quality effects on human health, crops and ecosystems
detection in industrial facilities,” IEEE Systems Journal, vol. 15, no. 4, change in the future?” Philosophical Transactions of the Royal Society
pp. 5135–5143, 2020. A, vol. 378, no. 2183, p. 20190330, 2020.
[49] A. Gaur, A. Singh, A. Kumar, K. S. Kulkarni, S. Lala, K. Kapoor, [68] D. Fowler, P. Brimblecombe, J. Burrows, M. R. Heal, P. Grennfelt,
V. Srivastava, A. Kumar, and S. C. Mukhopadhyay, “Fire sensing D. S. Stevenson, A. Jowett, E. Nemitz, M. Coyle, X. Liu et al., “A
technologies: A review,” IEEE Sensors Journal, vol. 19, no. 9, pp. chronology of global air quality,” Philosophical Transactions of the
3191–3202, 2019. Royal Society A, vol. 378, no. 2183, p. 20190314, 2020.
[50] C.-H. Huang, W.-T. Chen, Y.-C. Chang, and K.-T. Wu, “An edge [69] A. Baklanov and Y. Zhang, “Advances in air quality modeling and
and trustworthy AI UAV system with self-adaptivity and hyperspectral forecasting,” Global Transitions, vol. 2, pp. 261–270, 2020.
imaging for air quality monitoring,” IEEE Internet of Things Journal, [70] D. Singh, M. Dahiya, R. Kumar, and C. Nanda, “Sensors and systems
vol. 11, no. 20, p. 32572–32584, 2024. for air quality assessment monitoring and management: A review,”
[51] B. Arora, S. Kuppel, C. Wellen, C. Oswald, J. Groh, D. Payandi- Journal of Environmental Management, vol. 289, p. 112510, 2021.
Rolland, J. Stegen, and S. Coffinet, “Building cross-site and cross- [71] C. Silibello, A. Bolignano, R. Sozzi, and C. Gariazzo, “Application of
network collaborations in critical zone science,” Journal of Hydrology, a chemical transport model and optimized data assimilation methods
vol. 618, p. 129248, 2023. to improve air quality assessment,” Air Quality, Atmosphere & Health,
[52] N. Pandey, H. de Coninck, and A. D. Sagar, “Beyond technology vol. 7, pp. 283–296, 2014.
transfer: Innovation cooperation to advance sustainable development [72] A. Vlasenko, V. Matthias, and U. Callies, “Simulation of chemical
in developing countries,” Wiley Interdisciplinary Reviews: Energy and transport model estimates by means of a neural network using meteo-
Environment, vol. 11, no. 2, p. e422, 2022. rological data,” Atmospheric Environment, vol. 254, p. 118236, 2021.
[53] P. M. Pauley and B. Ingleby, “Assimilation of in-situ observations,” [73] C. Liu, H. Zhang, Z. Cheng, J. Shen, J. Zhao, Y. Wang, S. Wang, and
Data Assimilation for Atmospheric, Oceanic and Hydrologic Applica- Y. Cheng, “Emulation of an atmospheric gas-phase chemistry solver
tions (Vol. IV), pp. 293–371, 2022. through deep learning: Case study of Chinese mainland,” Atmospheric
[54] D. Bowdalo, S. Basart, M. Guevara, O. Jorba, C. Pérez Garcı́a- Pollution Research, vol. 12, no. 6, p. 101079, 2021.
Pando, M. Jaimes Palomera, O. Rivera Hernandez, M. Puchalski, [74] H. Liu, G. Yan, Z. Duan, and C. Chen, “Intelligent modeling strate-
D. Gay, J. Klausen, S. Moreno, S. Netcheva, and O. Tarasova, “Ghost: gies for forecasting air quality time series: A review,” Applied Soft
a globally harmonised dataset of surface atmospheric composition Computing, vol. 102, p. 106957, 2021.
measurements,” Earth System Science Data, vol. 16, no. 10, pp. 4417– [75] W. Leong, R. Kelani, and Z. Ahmad, “Prediction of air pollution index
4495, 2024. (API) using support vector machine (SVM),” Journal of Environmental
[55] C. Buontempo, S. N. Burgess, D. Dee, B. Pinty, J.-N. Thépaut, Chemical Engineering, vol. 8, no. 3, p. 103208, 2020.
M. Rixen, S. Almond, D. Armstrong, A. Brookshaw, A. L. Alos, [76] Y. Zhan, Y. Luo, X. Deng, M. L. Grieneisen, M. Zhang, and B. Di,
B. Bell, C. Bergeron, C. Cagnazzo, E. Comyn-Platt, E. Damasio-Da- “Spatiotemporal prediction of daily ambient ozone levels across China
Costa, A. Guillory, H. Hersbach, A. Horányi, J. Nicolas, A. Obregon, using random forest for human exposure assessment,” Environmental
E. P. Ramos, B. Raoult, J. Muñoz-Sabater, A. Simmons, C. Soci, Pollution, vol. 233, pp. 464–473, 2018.
 23

[77] R. Chauhan, H. Kaur, and B. Alankar, “Air quality forecast using variables: a mutual information approach,” Atmospheric Chemistry and
convolutional neural network for sustainable development in urban Physics, vol. 18, no. 17, pp. 12 699–12 714, 2018.
environments,” Sustainable Cities and Society, vol. 75, p. 103239, [96] P. Laarne, E. Amnell, M. A. Zaidan, S. Mikkonen, and T. Nieminen,
2021. “Exploring non-linear dependencies in atmospheric data with mutual
[78] N. Jin, Y. Zeng, K. Yan, and Z. Ji, “Multivariate air quality forecasting information,” Atmosphere, vol. 13, no. 7, p. 1046, 2022.
with nested long short term memory neural network,” IEEE Transac- [97] N. Djuric, L. Kansakar, and S. Vucetic, “Semi-supervised learning for
tions on Industrial Informatics, vol. 17, no. 12, pp. 8514–8522, 2021. integration of aerosol predictions from multiple satellite instruments,”
[79] J. Han, H. Liu, H. Xiong, and J. Yang, “Semi-supervised air quality in Proceedings of the Twenty-Third International Joint Conference
forecasting via self-supervised hierarchical graph neural network,” on Artificial Intelligence, ser. IJCAI ’13. AAAI Press, 2013, p.
IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 5, 2797–2803.
pp. 5230–5243, 2022. [98] M. L. Flora, C. K. Potvin, A. McGovern, and S. Handler, “A machine
[80] E. Austin, B. Coull, D. Thomas, and P. Koutrakis, “A framework learning explainability tutorial for atmospheric sciences,” Artificial
for identifying distinct multipollutant profiles in air pollution data,” Intelligence for the Earth Systems, vol. 3, no. 1, p. e230018, 2024.
Environment International, vol. 45, pp. 112–121, 2012. [99] V. Kukreja, T. P. S. Brar, S. Vats, A. Bansal, and R. Sharma, “Advanc-
[81] W. Huang, T. Li, J. Liu, P. Xie, S. Du, and F. Teng, “An overview ing climate forecasting accuracy through neurosymbolic integration:
of air quality analysis by big data techniques: Monitoring, forecasting, Unveiling the neurosymbolic climate inference and prediction (ns-cip)
and traceability,” Information Fusion, vol. 75, pp. 28–40, 2021. model’s approach to predicting extreme weather events,” SN Computer
[82] A. Borah, “Deep learning based anomaly detection approach for air Science, vol. 5, no. 7, p. 958, 2024.
pollution assessment,” IEEE Transactions on Big Data, 2024. [100] D. R. Harrison, A. McGovern, C. D. Karstens, A. Bostrom, J. L.
[83] L. Liang, J. Daniels, C. Bailey, L. Hu, R. Phillips, and J. South, Demuth, I. L. Jirak, and P. T. Marsh, “An assessment of how domain
“Integrating low-cost sensor monitoring, satellite mapping, and geospa- experts evaluate machine learning in operational meteorology,” Weather
tial artificial intelligence for intra-urban air pollution predictions,” and Forecasting, vol. 40, no. 3, pp. 393–410, 2025.
Environmental Pollution, vol. 331, p. 121832, 2023. [101] L. Chen, X. Zhong, H. Li, J. Wu, B. Lu, D. Chen, S.-P. Xie,
[84] T. Tasioulis and K. Karatzas, “Reviewing explainable artificial in- L. Wu, Q. Chao, C. Lin et al., “A machine learning model that
telligence towards better air quality modelling,” in Environmental outperforms conventional global subseasonal forecast models,” Nature
Informatics. Springer, 2023, pp. 3–19. Communications, vol. 15, no. 1, p. 6425, 2024.
[85] F. Concas, J. Mineraud, E. Lagerspetz, S. Varjonen, X. Liu, K. Puo- [102] M. G. Schultz, C. Betancourt, B. Gong, F. Kleinert, M. Langguth,
lamäki, P. Nurmi, and S. Tarkoma, “Low-cost outdoor air quality L. H. Leufen, A. Mozaffari, and S. Stadtler, “Can deep learning
monitoring and sensor calibration: A survey and critical analysis,” ACM beat numerical weather prediction?” Philosophical Transactions of the
Transactions on Sensor Networks (TOSN), vol. 17, no. 2, pp. 1–44, Royal Society A, vol. 379, no. 2194, p. 20200097, 2021.
2021. [103] K. Bi, L. Xie, H. Zhang, X. Chen, X. Gu, and Q. Tian, “Accurate
[86] L. Fu, J. Li, and Y. Chen, “An innovative decision making method for medium-range global weather forecasting with 3d neural networks,”
air quality monitoring based on big data-assisted artificial intelligence Nature, vol. 619, no. 7970, pp. 533–538, 2023.
technique,” Journal of Innovation & Knowledge, vol. 8, no. 2, p. [104] J. Wei, M. Chen, L. Wang, P. Ren, Y. Lei, Y. Qu, Q. Jiang, X. Dong,
100294, 2023. W. Wu, Q. Wang et al., “Status, challenges and trends of data-intensive
[87] M. Xu, J. Jin, G. Wang, A. Segers, T. Deng, and H. X. Lin, “Ma- supercomputing,” CCF Transactions on High Performance Computing,
chine learning based bias correction for numerical chemical transport vol. 4, no. 2, pp. 211–230, 2022.
models,” Atmospheric Environment, vol. 248, p. 118022, 2021. [105] L. Espeholt, S. Agrawal, C. Sønderby, M. Kumar, J. Heek,
[88] R. Philipona, “Networks of atmospheric measuring techniques,” C. Bromberg, C. Gazen, R. Carver, M. Andrychowicz, J. Hickey
Springer Handbook of Atmospheric Measurements, pp. 1677–1698, et al., “Deep learning for twelve hour precipitation forecasts,” Nature
2021. communications, vol. 13, no. 1, pp. 1–10, 2022.
[89] M. Kulmala, T. Petäjä, T. Nieminen, M. Sipilä, H. E. Manninen, [106] K. K. Jena, S. K. Bhoi, S. R. Nayak, R. Panigrahi, and A. K. Bhoi,
K. Lehtipalo, M. Dal Maso, P. P. Aalto, H. Junninen, P. Paasonen, “Deep convolutional network based machine intelligence model for
I. Riipinen, K. E. J. Lehtinen, A. Laaksonen, and V.-M. Kerminen, satellite cloud image classification,” Big Data Mining and Analytics,
“Measurement of the nucleation of atmospheric aerosol particles,” vol. 6, no. 1, pp. 32–43, 2022.
Nature Protocols, vol. 7, no. 9, pp. 1651–1667, 2012. [107] C. Wang, G. Zheng, X. Li, Q. Xu, B. Liu, and J. Zhang, “Tropical
[90] M. Kulmala, J. Kontkanen, H. Junninen, K. Lehtipalo, H. E. Manninen, cyclone intensity estimation from geostationary satellite imagery using
T. Nieminen, T. Petäjä, M. Sipilä, S. Schobesberger, P. Rantala, deep convolutional neural networks,” IEEE Transactions on Geoscience
A. Franchin, T. Jokinen, E. Järvinen, M. Äijälä, J. Kangasluoma, and Remote Sensing, vol. 60, pp. 1–16, 2021.
J. Hakala, P. P. Aalto, P. Paasonen, J. Mikkilä, J. Vanhanen, J. Aalto, [108] M. A. R. Suleman and S. Shridevi, “Short-term weather forecasting
H. Hakola, U. Makkonen, T. Ruuskanen, R. L. Mauldin, J. Duplissy, using spatial feature attention based lstm model,” IEEE Access, vol. 10,
H. Vehkamäki, J. Bäck, A. Kortelainen, I. Riipinen, T. Kurtén, M. V. pp. 82 456–82 468, 2022.
Johnston, J. N. Smith, M. Ehn, T. F. Mentel, K. E. J. Lehtinen, A. Laak- [109] P. Grönquist, C. Yao, T. Ben-Nun, N. Dryden, P. Dueben, S. Li,
sonen, V.-M. Kerminen, and D. R. Worsnop, “Direct observations of and T. Hoefler, “Deep learning for post-processing ensemble weather
atmospheric aerosol nucleation,” Science, vol. 339, no. 6122, pp. 943– forecasts,” Philosophical Transactions of the Royal Society A, vol. 379,
946, 2013. no. 2194, p. 20200092, 2021.
[91] M. Dal Maso, M. Kulmala, I. Riipinen, R. Wagner, T. Hussein, P. P. [110] A. Bihlo, “A generative adversarial network approach to (ensemble)
Aalto, and K. E. Lehtinen, “Formation and growth of fresh atmospheric weather prediction,” Neural Networks, vol. 139, pp. 1–16, 2021.
aerosols: eight years of aerosol size distribution data from SMEAR II, [111] S. Bojinski, D. Blaauboer, X. Calbet, E. De Coning, F. Debie, T. Mont-
hyytiälä, finland,” Boreal Environment Research, vol. 10, no. 5, p. 323, merle, V. Nietosvaara, K. Norman, L. Bañón Peregrı́n, F. Schmid et al.,
2005. “Towards nowcasting in europe in 2030,” Meteorological applications,
[92] M. Zaidan, V. Haapasilta, R. Relan, H. Junninen, P. Aalto, M. Kulmala, vol. 30, no. 4, p. e2124, 2023.
L. Laurson, and A. Foster, “Predicting atmospheric particle formation [112] C. Feng, M. Cui, B.-M. Hodge, S. Lu, H. F. Hamann, and J. Zhang,
days by bayesian classification of the time series features,” Tellus B: “Unsupervised clustering-based short-term solar forecasting,” IEEE
Chemical and Physical Meteorology, vol. 70, no. 1, pp. 1–10, 2018. Transactions on Sustainable Energy, vol. 10, no. 4, pp. 2174–2185,
[93] P. Su, J. Joutsensaari, L. Dada, M. A. Zaidan, T. Nieminen, X. Li, 2018.
Y. Wu, S. Decesari, S. Tarkoma, T. Petäjä, M. Kulmala, and P. Pel- [113] R. Yang, J. Hu, Z. Li, J. Mu, T. Yu, J. Xia, X. Li, A. Dasgupta,
likka, “New particle formation event detection with Mask R-CNN,” and H. Xiong, “Interpretable machine learning for weather and climate
Atmospheric Chemistry and Physics, vol. 22, no. 2, pp. 1293–1309, prediction: A review,” Atmospheric Environment, p. 120797, 2024.
2022. [114] J.-P. Montillet, G. Kermarrec, E. Forootan, M. Haberreiter, X. He,
[94] A. Marinoni and P. Gamba, “Unsupervised data driven feature extrac- W. Finsterle, R. Fernandes, and C. Shum, “How big data can help to
tion by means of mutual information maximization,” IEEE Transac- monitor the environment and to mitigate risks due to climate change:
tions on Computational Imaging, vol. 3, no. 2, pp. 243–253, 2017. a review,” IEEE Geoscience and Remote Sensing Magazine, 2024.
[95] M. A. Zaidan, V. Haapasilta, R. Relan, P. Paasonen, V.-M. Kerminen, [115] D. Kochkov, J. Yuval, I. Langmore, P. Norgaard, J. Smith, G. Mooers,
H. Junninen, M. Kulmala, and A. S. Foster, “Exploring non-linear M. Klöwer, J. Lottes, S. Rasp, P. Düben et al., “Neural general
associations between atmospheric new-particle formation and ambient circulation models for weather and climate,” Nature, pp. 1–7, 2024.
 24

[116] F. F. Sabins Jr and J. M. Ellis, Remote sensing: Principles, interpreta- learning in earth observation,” IEEE Geoscience and Remote Sensing
tion, and applications. Waveland Press, 2020. Magazine, vol. 11, no. 3, pp. 63–97, 2023.
[117] J. S. Fu, G. R. Carmichael, F. Dentener, W. Aas, C. Andersson, [134] S. Wang, W. Han, X. Huang, X. Zhang, L. Wang, and J. Li, “Trust-
L. A. Barrie, A. Cole, C. Galy-Lacaux, J. Geddes, S. Itahashi et al., worthy remote sensing interpretation: Concepts, technologies, and
“Improving estimates of sulfur, nitrogen, and ozone total deposition applications,” ISPRS Journal of Photogrammetry and Remote Sensing,
through multi-model and measurement-model fusion approaches,” En- vol. 209, pp. 150–172, 2024.
vironmental science & technology, vol. 56, no. 4, pp. 2134–2142, 2022. [135] G. B. Bonan and S. C. Doney, “Climate, ecosystems, and planetary
[118] J. D. Shutler, N. Gruber, H. S. Findlay, P. E. Land, L. Gregor, futures: The challenge to predict life in earth system models,” Science,
T. Holding, R. P. Sims, H. Green, J.-F. Piolle, B. Chapron et al., vol. 359, no. 6375, Feb. 2018.
“The increasing importance of satellite observations to assess the ocean [136] A. Gettelman, A. J. Geer, R. M. Forbes, G. R. Carmichael, G. Feingold,
carbon sink and ocean acidification,” Earth-Science Reviews, vol. 250, D. J. Posselt, G. L. Stephens, S. C. Van den Heever, A. C. Varble,
p. 104682, 2024. and P. Zuidema, “The future of earth system prediction: Advances in
[119] R. Avtar, A. A. Komolafe, A. Kouser, D. Singh, A. P. Yunus, J. Dou, model-data fusion,” Science Advances, vol. 8, no. 14, p. eabn3488,
P. Kumar, R. D. Gupta, B. A. Johnson, H. V. T. Minh et al., “Assessing 2022.
sustainable development prospects through remote sensing: A review,” [137] K. E. Trenberth, J. T. Fasullo, and T. G. Shepherd, “Attribution of
Remote sensing applications: Society and environment, vol. 20, p. climate extreme events,” Nature Climate Change, vol. 5, no. 8, p.
100402, 2020. 725–730, Jun. 2015.
[120] B. Zhang, Y. Wu, B. Zhao, J. Chanussot, D. Hong, J. Yao, and L. Gao, [138] C. D. Hewitt, F. Guglielmo, S. Joussaume, J. Bessembinder, I. Christel,
“Progress and challenges in intelligent remote sensing satellite sys- F. J. Doblas-Reyes, V. Djurdjevic, N. Garrett, E. Kjellström, A. Krzic,
tems,” IEEE Journal of Selected Topics in Applied Earth Observations M. M. Costa, and A. L. St. Clair, “Recommendations for future research
and Remote Sensing, vol. 15, pp. 1814–1822, 2022. priorities for climate modeling and climate services,” Bulletin of the
[121] L. Zhang and L. Zhang, “Artificial intelligence for remote sensing data American Meteorological Society, vol. 102, no. 3, p. E578–E588, Mar.
analysis: A review of challenges and opportunities,” IEEE Geoscience 2021.
and Remote Sensing Magazine, vol. 10, no. 2, pp. 270–294, 2022. [139] C. Irrgang, N. Boers, M. Sonnewald, E. A. Barnes, C. Kadow,
[122] W. Han, X. Zhang, Y. Wang, L. Wang, X. Huang, J. Li, S. Wang, J. Staneva, and J. Saynisch-Wagner, “Towards neural earth system
W. Chen, X. Li, R. Feng et al., “A survey of machine learning and modelling by integrating artificial intelligence in earth system science,”
deep learning in remote sensing of geological environment: Challenges, Nature Machine Intelligence, vol. 3, no. 8, p. 667–674, Aug. 2021.
advances, and opportunities,” ISPRS Journal of Photogrammetry and [140] M. Reichstein, G. Camps-Valls, B. Stevens, M. Jung, J. Denzler,
Remote Sensing, vol. 202, pp. 87–113, 2023. N. Carvalhais, and F. Prabhat, “Deep learning and process understand-
[123] C. Persello, J. D. Wegner, R. Hänsch, D. Tuia, P. Ghamisi, M. Ko- ing for data-driven earth system science,” Nature, vol. 566, no. 7743,
eva, and G. Camps-Valls, “Deep learning and earth observation to pp. 195–204, 2019.
support the sustainable development goals: Current approaches, open [141] S. Rasp, M. S. Pritchard, and P. Gentine, “Deep learning to represent
challenges, and future opportunities,” IEEE Geoscience and Remote subgrid processes in climate models,” Proceedings of the national
Sensing Magazine, vol. 10, no. 2, pp. 172–200, 2022. academy of sciences, vol. 115, no. 39, pp. 9684–9689, 2018.
[142] Y.-G. Ham, J.-H. Kim, and J.-J. Luo, “Deep learning for multi-year
[124] L. Miller, C. Pelletier, and G. I. Webb, “Deep learning for satellite
enso forecasts,” Nature, vol. 573, no. 7775, pp. 568–572, 2019.
image time-series analysis: A review,” IEEE Geoscience and Remote
[143] S. Hoffmann and C. Lessig, “Atmodist: Self-supervised representation
Sensing Magazine, 2024.
learning for atmospheric dynamics,” Environmental Data Science,
[125] Y. Cao, Y. Wang, J. Peng, C. Qiu, L. Ding, and X. X. Zhu, “Sdfl-
vol. 2, p. e6, 2023.
fc: Semisupervised deep feature learning with feature consistency for
[144] G. Mooers, M. Pritchard, T. Beucler, P. Srivastava, H. Mangipudi,
hyperspectral image classification,” IEEE Transactions on Geoscience
L. Peng, P. Gentine, and S. Mandt, “Comparing storm resolving models
and Remote Sensing, vol. 59, no. 12, pp. 10 488–10 502, 2020.
and climates via unsupervised machine learning,” Scientific Reports,
[126] Y. Wang, C. M. Albrecht, N. A. A. Braham, L. Mou, and X. X. vol. 13, no. 1, p. 22365, 2023.
Zhu, “Self-supervised learning in remote sensing: A review,” IEEE [145] Z. Li, “Geospatial big data handling with high performance computing:
Geoscience and Remote Sensing Magazine, vol. 10, no. 4, pp. 213– Current approaches and future directions,” High Performance Comput-
247, 2022. ing for Geospatial Applications, pp. 53–76, 2020.
[127] J. Guo, J. Yang, H. Yue, X. Liu, and K. Li, “Unsupervised domain- [146] P. Mazzoglio, E. Danovaro, L. Ganne, A. Parodi, S. Hachinger,
invariant feature learning for cloud detection of remote sensing im- A. Galizia, A. Parodi, and J. Martinovič, “Exploitation of multiple
ages,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, model layers within lexis weather and climate pilot: An hpc-based
pp. 1–15, 2021. approach,” in HPC, Big Data, and AI Convergence Towards Exascale.
[128] P. Wang, B. Bayram, and E. Sertel, “A comprehensive review on deep CRC Press, 2022, pp. 146–163.
learning based remote sensing image super-resolution methods,” Earth- [147] T. de Wolff, H. Carrillo, L. Martı́, and N. Sanchez-Pi, “Assessing
Science Reviews, vol. 232, p. 104110, 2022. physics informed neural networks in ocean modelling and climate
[129] M. Sdraka, I. Papoutsis, B. Psomas, K. Vlachos, K. Ioannidis, change applications,” in AI: Modeling Oceans and Climate Change
K. Karantzalos, I. Gialampoukidis, and S. Vrochidis, “Deep learning Workshop at ICLR 2021, 2021.
for downscaling remote sensing images: Fusion and super-resolution,” [148] A. Meray, L. Wang, T. Kurihana, I. Mastilovic, S. Praveen, Z. Xu,
IEEE Geoscience and Remote Sensing Magazine, vol. 10, no. 3, pp. M. Memarzadeh, A. Lavin, and H. Wainwright, “Physics-informed
202–255, 2022. surrogate modeling for supporting climate resilience at groundwater
[130] Y. Himeur, B. Rimal, A. Tiwary, and A. Amira, “Using artificial contamination sites,” Computers & Geosciences, vol. 183, p. 105508,
intelligence and data fusion for environmental monitoring: A review 2024.
and future perspectives,” Information Fusion, vol. 86–87, p. 44–75, [149] G. B. Bonan and S. C. Doney, “Climate, ecosystems, and planetary
Oct. 2022. futures: The challenge to predict life in earth system models,” Science,
[131] T. Zhao, S. Wang, C. Ouyang, M. Chen, C. Liu, J. Zhang, L. Yu, vol. 359, no. 6375, p. eaam8328, 2018.
F. Wang, Y. Xie, J. Li, F. Wang, S. Grunwald, B. M. Wong, F. Zhang, [150] R. Lamsal and T. V. Kumar, “Artificial intelligence and early warning
Z. Qian, Y. Xu, C. Yu, W. Han, T. Sun, Z. Shao, T. Qian, Z. Chen, systems,” AI and Robotics in Disaster Studies, pp. 13–32, 2020.
J. Zeng, H. Zhang, H. Letu, B. Zhang, L. Wang, L. Luo, C. Shi, [151] A. Albahri, Y. L. Khaleel, M. A. Habeeb, R. D. Ismael, Q. A. Hameed,
H. Su, H. Zhang, S. Yin, N. Huang, W. Zhao, N. Li, C. Zheng, M. Deveci, R. Z. Homod, O. Albahri, A. Alamoodi, and L. Alzubaidi,
Y. Zhou, C. Huang, D. Feng, Q. Xu, Y. Wu, D. Hong, Z. Wang, Y. Lin, “A systematic review of trustworthy artificial intelligence applications
T. Zhang, P. Kumar, A. Plaza, J. Chanussot, J. Zhang, J. Shi, and in natural disasters,” Computers and Electrical Engineering, vol. 118,
L. Wang, “Artificial intelligence for geoscience: Progress, challenges, p. 109409, 2024.
and perspectives,” The Innovation, vol. 5, no. 5, p. 100691, Sep. 2024. [152] W.-j. Niu and Z.-k. Feng, “Evaluating the performances of several
[132] C. Xu, X. Du, X. Fan, G. Giuliani, Z. Hu, W. Wang, J. Liu, T. Wang, artificial intelligence methods in forecasting daily streamflow time
Z. Yan, J. Zhu et al., “Cloud-based storage and computing for remote series for sustainable water resources management,” Sustainable Cities
sensing big data: a technical review,” International Journal of Digital and Society, vol. 64, p. 102562, 2021.
Earth, vol. 15, no. 1, pp. 1417–1445, 2022. [153] F.-J. Chang, L.-C. Chang, and J.-F. Chen, “Artificial intelligence
[133] M. Schmitt, S. A. Ahmadi, Y. Xu, G. Taşkin, U. Verma, F. Sica, techniques in hydrology and water resources management,” p. 1846,
and R. Hänsch, “There are no data like more data: Datasets for deep 2023.
 25

[154] A. Sharma, A. Jain, P. Gupta, and V. Chowdary, “Machine learning lenges and opportunities of near-term quantum computing systems,”
applications for precision agriculture: A comprehensive review,” IEEE Proceedings of the IEEE, vol. 108, no. 8, pp. 1338–1352, 2019.
Access, vol. 9, pp. 4843–4873, 2020. [174] S. Ashwani, A. J. Tripathy, S. Karna, P. R. Jahanve, and S. M.
[155] A. Leonardi, H. M. Ricker, A. G. Gale, B. T. Ball, T. T. Odbadrakh, Rajagopal, “Quantum computing for climate change: A comprehensive
G. C. Shields, and J. G. Navea, “Particle formation and surface review of current applications, challenges, and future directions,” in
processes on atmospheric aerosols: A review of applied quantum 2024 15th International Conference on Computing Communication and
chemical calculations,” International Journal of Quantum Chemistry, Networking Technologies (ICCCNT). IEEE, 2024, pp. 1–7.
vol. 120, no. 20, p. e26350, 2020. [175] A. Frolov, “Can a quantum computer be applied for numerical weather
[156] S. Javaid, H. Fahim, S. Zeadally, and B. He, “Self-powered sensors: Ap- prediction?” Russian meteorology and hydrology, vol. 42, pp. 545–553,
plications, challenges, and solutions,” IEEE Sensors Journal, vol. 23, 2017.
no. 18, pp. 20 483–20 509, 2023. [176] S. Suhas and S. Divya, “Quantum-improved weather forecasting: Inte-
[157] J. Ahn, S. Cho, L. Wu, X. Li, D. Lee, J.-H. Ha, H. Han, K. Lee, grating quantum machine learning for precise prediction and disaster
B. Kang, Y. Kwon et al., “Innovations in self-powered sensors utilizing mitigation,” in 2023 International Conference on Quantum Technolo-
light, thermal, and mechanical renewable energy,” Nano Energy, p. gies, Communications, Computing, Hardware and Embedded Systems
110045, 2024. Security (iQ-CCHESS). IEEE, 2023, pp. 1–7.
[158] R. M. Ferdous, A. W. Reza, and M. F. Siddiqui, “Renewable energy [177] R. Y. Patil, Y. H. Patil, and S. Doss, “Utilizing quantum computing for
harvesting for wireless sensors using passive rfid tag technology: A enhanced natural disaster prediction and mitigation strategies,” in The
review,” Renewable and Sustainable Energy Reviews, vol. 58, pp. 1114– Rise of Quantum Computing in Industry 6.0 Towards Sustainability.
1128, 2016. Springer, 2024, pp. 141–154.
[159] F. d. L. L. de Amorim, K. H. Wiltshire, P. Lemke, K. Carstens, [178] H. T. Nguyen, M. Usman, and R. Buyya, “Qfaas: A serverless function-
S. Peters, J. Rick, L. Gimenez, and M. Scharfe, “Investigation of marine as-a-service framework for quantum computing,” Future Generation
temperature changes across temporal and spatial gradients: providing a Computer Systems, vol. 154, pp. 281–300, 2024.
fundament for studies on the effects of warming on marine ecosystem [179] T. Luo, W.-F. Wong, R. S. M. Goh, A. T. Do, Z. Chen, H. Li, W. Jiang,
function and biodiversity,” Progress in Oceanography, vol. 216, p. and W. Yau, “Achieving green AI with energy-efficient deep learning
103080, 2023. using neuromorphic computing,” Communications of the ACM, vol. 66,
[160] P. Ngoy, F. Dar, M. Liyanage, Z. Yin, U. Norbisrath, A. Zuniga, no. 7, pp. 52–57, 2023.
J. Pestana, M. Radeta, P. Nurmi, and H. Flores, “Supporting sustainable [180] R. Li, Y. Gong, H. Huang, Y. Zhou, S. Mao, Z. Wei, and Z. Zhang,
computing by repurposing e-waste smartphones as tiny data centres,” “Photonics for neuromorphic computing: Fundamentals, devices, and
IEEE Pervasive Computing, 2025. opportunities,” Advanced Materials, vol. n/a, no. n/a, p. 2312825, 2024.
[161] N. Fumeaux, M. Kossairi, J. Bourely, and D. Briand, “Printed ecore-
[181] M. Davies, A. Wild, G. Orchard, Y. Sandamirskaya, G. A. F. Guerra,
sorbable temperature sensors for environmental monitoring,” Micro and
P. Joshi, P. Plank, and S. R. Risbud, “Advancing neuromorphic com-
Nano Engineering, vol. 20, p. 100218, 2023.
puting with loihi: A survey of results and outlook,” Proceedings of the
[162] A. Zahedi, R. Liyanapathirana, and K. Thiyagarajan, “Biodegradable IEEE, vol. 109, no. 5, pp. 911–934, 2021.
and renewable antennas for green iot sensors: A review,” IEEE Access,
[182] C. Bodnar, W. P. Bruinsma, A. Lucic, M. Stanley, J. Brandstetter,
pp. 1–1, 2024.
P. Garvan, M. Riechert, J. Weyn, H. Dong, A. Vaughan et al.,
[163] M. A. A. Mamun and M. R. Yuce, “Recent progress in nanomaterial
“Aurora: A foundation model of the atmosphere,” arXiv preprint
enabled chemical sensors for wearable environmental monitoring appli-
arXiv:2405.13063, 2024.
cations,” Advanced Functional Materials, vol. 30, no. 51, p. 2005703,
2020. [183] I. Price, A. Sanchez-Gonzalez, F. Alet, T. R. Andersson, A. El-Kadi,
D. Masters, T. Ewalds, J. Stott, S. Mohamed, P. Battaglia et al.,
[164] M. M. Fouda, S. Sakib, Z. M. Fadlullah, N. Nasser, and M. Guizani,
“Probabilistic weather forecasting with machine learning,” Nature, pp.
“A lightweight hierarchical ai model for uav-enabled edge computing
1–7, 2024.
with forest-fire detection use-case,” IEEE Network, vol. 36, no. 6, pp.
38–45, 2022. [184] C. Vineyard, W. Severa, M. Kagie, A. Scholand, and P. Hays, “A
[165] A. Nikul Patel, G. Srivastava, P. Kumar Reddy Maddikunta, R. Mu- resurgence in neuromorphic architectures enabling remote sensing
rugan, G. Yenduri, and T. Reddy Gadekallu, “A trustable federated computation,” in 2019 IEEE Space Computing Conference (SCC).
learning framework for rapid fire smoke detection at the edge in smart IEEE, 2019, pp. 33–40.
home environments,” IEEE Internet of Things Journal, vol. 11, no. 23, [185] L. P. Garcia, G. Furano, M. Ghiglione, V. Zancan, E. Imbembo,
pp. 37 708–37 717, 2024. C. Ilioudis, C. Clemente, and P. Trucco, “Advancements in on-board
[166] M. A. Zaidan, N. H. Motlagh, P. L. Fung, A. S. Khalaf, Y. Matsumi, processing of synthetic aperture radar (sar) data: enhancing efficiency
A. Ding, S. Tarkoma, T. Petäjä, M. Kulmala, and T. Hussein, “Intel- and real-time capabilities,” IEEE Journal of Selected Topics in Applied
ligent air pollution sensors calibration for extreme events and drifts Earth Observations and Remote Sensing, 2024.
monitoring,” IEEE Transactions on Industrial Informatics, vol. 19, [186] Q. Liu, K. Yang, J. Xie, and Y. Sun, “Dna-based molecular computing,
no. 2, pp. 1366–1379, 2023. storage, and communications,” IEEE Internet of Things Journal, vol. 9,
[167] S. E. Bibri, J. Krogstie, A. Kaboli, and A. Alahi, “Smarter eco- no. 2, pp. 897–915, 2021.
cities and their leading-edge artificial intelligence of things solutions [187] S. Yang, B. W. Bögels, F. Wang, C. Xu, H. Dou, S. Mann, C. Fan, and
for environmental sustainability: A comprehensive systematic review,” T. F. de Greef, “Dna as a universal chemical substrate for computing
Environmental Science and Ecotechnology, vol. 19, p. 100330, 2024. and data storage,” Nature Reviews Chemistry, vol. 8, no. 3, pp. 179–
[168] S. Abulgasem, F. Tubbal, R. Raad, P. I. Theoharis, S. Lu, and 194, 2024.
S. Iranmanesh, “Antenna designs for cubesats: A review,” IEEE Access, [188] H. Jiao, Y. Zhong, and L. Zhang, “Artificial dna computing-based
vol. 9, pp. 45 289–45 324, 2021. spectral encoding and matching algorithm for hyperspectral remote
[169] N. Saeed, A. Elzanaty, H. Almorad, H. Dahrouj, T. Y. Al-Naffouri, and sensing data,” IEEE Transactions on Geoscience and Remote Sensing,
M.-S. Alouini, “Cubesat communications: Recent advances and future vol. 50, no. 10, pp. 4085–4104, 2012.
challenges,” IEEE Communications Surveys & Tutorials, vol. 22, no. 3, [189] E. T. M. Beltrán, M. Q. Pérez, P. M. S. Sánchez, S. L. Bernal,
pp. 1839–1862, 2020. G. Bovet, M. G. Pérez, G. M. Pérez, and A. H. Celdrán, “Decentralized
[170] M. C. Baddock, R. G. Bryant, M. D. Acosta, and T. E. Gill, “Un- federated learning: Fundamentals, state of the art, frameworks, trends,
derstanding dust sources through remote sensing: Making a case for and challenges,” IEEE Communications Surveys & Tutorials, 2023.
cubesats,” Journal of Arid Environments, vol. 184, p. 104335, 2021. [190] C. G. De Vita, G. Mellone, A. Casolaro, M. G. Orsini, J. L. Gonzalez-
[171] A. Gregorio and F. Alimenti, “Cubesats for future science and internet Compean, and A. Ciaramella, “Federated learning and crowdsourced
of space: Challenges and opportunities,” in Proceedings of the IEEE In- weather data: Practice and experience,” in 2024 IEEE 20th Interna-
ternational Conference on Electronics, Circuits, and Systems (ICECS), tional Conference on e-Science (e-Science). IEEE, 2024, pp. 1–9.
Dec. 2018, pp. 169–172. [191] J. Xu, W. Du, Q. Xu, J. Dong, and B. Wang, “Federated learning based
[172] S. K. Sood et al., “Quantum computing review: A decade of research,” atmospheric source term estimation in urban environments,” Computers
IEEE Transactions on Engineering Management, vol. 71, pp. 6662– & Chemical Engineering, vol. 155, p. 107505, 2021.
6676, 2023. [192] S. Moreno-Álvarez, M. E. Paoletti, A. J. Sanchez-Fernandez, J. A.
[173] A. D. Córcoles, A. Kandala, A. Javadi-Abhari, D. T. McClure, A. W. Rico-Gallego, L. Han, and J. M. Haut, “Federated learning meets
Cross, K. Temme, P. D. Nation, M. Steffen, and J. M. Gambetta, “Chal- remote sensing,” Expert Systems with Applications, p. 124583, 2024.
 26

[193] X. Qiu, T. Parcollet, D. J. Beutel, T. Topal, A. Mathur, and N. D. Lane, [213] S. Li and T. Wu, “A risk-informed decision-support framework for opti-
“Can federated learning save the planet?” in NeurIPS-Tackling Climate mal operation of hurricane-impacted transportation networks,” Natural
Change with Machine Learning, 2020. hazards review, vol. 24, no. 3, p. 04023018, 2023.
[194] M. Ye, X. Fang, B. Du, P. C. Yuen, and D. Tao, “Heterogeneous [214] W. Hazeleger, J. Aerts, P. Bauer, M. Bierkens, G. Camps-Valls,
federated learning: State-of-the-art and research challenges,” ACM M. Dekker, F. Doblas-Reyes, V. Eyring, C. Finkenauer, A. Grundner
Computing Surveys, vol. 56, no. 3, pp. 1–44, 2023. et al., “Digital twins of the earth with and for humans,” Communica-
[195] K. Li, X. Cao, and D. Meng, “A new learning paradigm for foundation tions Earth & Environment, vol. 5, no. 1, p. 463, 2024.
model-based remote-sensing change detection,” IEEE Transactions on [215] T. Hoefler, B. Stevens, A. F. Prein, J. Baehr, T. Schulthess, T. F.
Geoscience and Remote Sensing, vol. 62, pp. 1–12, 2024. Stocker, J. Taylor, D. Klocke, P. Manninen, P. M. Forster et al., “Earth
[196] D. Hong, B. Zhang, X. Li, Y. Li, C. Li, J. Yao, N. Yokoya, H. Li, virtualization engines: a technical perspective,” Computing in Science
P. Ghamisi, X. Jia et al., “SpectralGPT: Spectral remote sensing foun- & Engineering, vol. 25, no. 3, pp. 50–59, 2023.
dation model,” IEEE Transactions on Pattern Analysis and Machine [216] B. Stevens, S. Adami, T. Ali, H. Anzt, Z. Aslan, S. Attinger, J. Bäck,
Intelligence, 2024. J. Baehr, P. Bauer, N. Bernier et al., “Earth virtualization engines (eve),”
[197] Z. Lin, C. Deng, L. Zhou, T. Zhang, Y. Xu, Y. Xu, Z. He, Y. Shi, Earth System Science Data, vol. 16, no. 4, pp. 2113–2122, 2024.
B. Dai, Y. Song et al., “Geogalactica: A scientific large language model [217] P. Bauer, B. Stevens, and W. Hazeleger, “A digital twin of earth for the
in geoscience,” arXiv preprint arXiv:2401.00434, 2023. green transition,” Nature Climate Change, vol. 11, pp. 80–83, 2021.
[198] D. H. Hagos, R. Battle, and D. B. Rawat, “Recent advances in [218] D. Peldon, S. Banihashemi, K. LeNguyen, and S. Derrible, “Navigating
generative ai and large language models: Current status, challenges, urban complexity: The transformative role of digital twins in smart
and perspectives,” IEEE Transactions on Artificial Intelligence, 2024. city development,” Sustainable Cities and Society, vol. 111, p. 105583,
[199] S. Bengesi, H. El-Sayed, M. K. Sarker, Y. Houkpati, J. Irungu, and 2024.
T. Oladunni, “Advancements in generative ai: A comprehensive review [219] G. Senka, M. Tramonti, A. M. Dochshanov, T. Jesmin, J. Terasmaa,
of gans, gpt, autoencoders, diffusion model, and transformers.” IEEe H. Tsalapatas, O. Heidmann, M. Caeiro-Rodriguez, and C. Vaz de
Access, 2024. Carvalho, “Using a game to educate about sustainable development,”
[200] R. Dwivedi, D. Dave, H. Naik, S. Singhal, R. Omer, P. Patel, B. Qian, Multimodal Technologies and Interaction, vol. 8, no. 11, p. 96, 2024.
Z. Wen, T. Shah, G. Morgan et al., “Explainable ai (xai): Core ideas, [220] P. G. R. de Almeida, C. D. dos Santos, and J. S. Farias, “Artificial
techniques, and solutions,” ACM Computing Surveys, vol. 55, no. 9, Intelligence regulation: a framework for governance,” Ethics and In-
pp. 1–33, 2023. formation Technology, vol. 23, no. 3, pp. 505–525, 2021.
[201] K. Jeggle, D. Neubauer, G. Camps-Valls, and U. Lohmann, “Under- [221] M. Janssen, P. Brous, E. Estevez, L. S. Barbosa, and T. Janowski, “Data
standing cirrus clouds using explainable machine learning,” Environ- governance: Organizing data for trustworthy Artificial Intelligence,”
mental Data Science, vol. 2, p. e19, 2023. Government information quarterly, vol. 37, no. 3, p. 101493, 2020.
[222] R. Joshi, K. Pandey, and S. Kumari, “Artificial intelligence for advanced
[202] P. L. Bommer, M. Kretschmer, A. Hedström, D. Bareeva, and M. M.-
sustainable development goals: A 360-degree approach,” in Preserv-
C. Höhne, “Finding the right XAI method—a guide for the evaluation
ing Health, Preserving Earth: The Path to Sustainable Healthcare.
and ranking of explainable AI methods in climate science,” Artificial
Springer, 2024, pp. 281–303.
Intelligence for the Earth Systems, vol. 3, no. 3, p. e230074, 2024.
[223] I. Kulkov, J. Kulkova, R. Rohrbeck, L. Menvielle, V. Kaartemo, and
[203] A. Fania, A. Monaco, N. Amoroso, L. Bellantuono, R. Cazzolla Gatti,
H. Makkonen, “Artificial intelligence-driven sustainable development:
N. Firza, A. Lacalamita, E. Pantaleo, S. Tangaro, A. Velichevskaya
Examining organizational, technical, and processing approaches to
et al., “Machine learning and xai approaches highlight the strong
achieving global goals,” Sustainable Development, vol. 32, no. 3, pp.
connection between o 3 and no 2 pollutants and alzheimer’s disease,”
2253–2267, 2024.
Scientific Reports, vol. 14, no. 1, p. 5385, 2024.
[224] N. Dı́az-Rodrı́guez, J. Del Ser, M. Coeckelbergh, M. L. de Prado,
[204] G. Carleo, I. Cirac, K. Cranmer, L. Daudet, M. Schuld, N. Tishby, E. Herrera-Viedma, and F. Herrera, “Connecting the dots in trustworthy
L. Vogt-Maranto, and L. Zdeborová, “Machine learning and the phys- artificial intelligence: From AI principles, ethics, and key requirements
ical sciences,” Reviews of Modern Physics, vol. 91, no. 4, p. 045002, to responsible ai systems and regulation,” Information Fusion, vol. 99,
2019. p. 101896, 2023.
[205] A. Karpatne, G. Atluri, J. H. Faghmous, M. Steinbach, A. Banerjee, [225] Y. K. Dwivedi, L. Hughes, E. Ismagilova, G. Aarts, C. Coombs,
A. Ganguly, S. Shekhar, N. Samatova, and V. Kumar, “Theory-guided T. Crick, Y. Duan, R. Dwivedi, J. Edwards, A. Eirug et al., “Ar-
data science: A new paradigm for scientific discovery from data,” IEEE tificial intelligence (ai): Multidisciplinary perspectives on emerging
Transactions on knowledge and data engineering, vol. 29, no. 10, pp. challenges, opportunities, and agenda for research, practice and policy,”
2318–2331, 2017. International Journal of Information Management, vol. 57, p. 101994,
[206] S. Cuomo, V. S. Di Cola, F. Giampaolo, G. Rozza, M. Raissi, and 2021.
F. Piccialli, “Scientific machine learning through physics–informed
neural networks: Where we are and what’s next,” Journal of Scientific
Computing, vol. 92, no. 3, p. 88, 2022. B IOGRAPHIES
[207] K. Kashinath, M. Mustafa, A. Albert, J. Wu, C. Jiang, S. Esmaeilzadeh,
K. Azizzadenesheli, R. Wang, A. Chattopadhyay, A. Singh et al.,
“Physics-informed machine learning: case studies for weather and
climate modelling,” Philosophical Transactions of the Royal Society
A, vol. 379, no. 2194, p. 20200093, 2021.
[208] T. Nguyen, J. Brandstetter, A. Kapoor, J. K. Gupta, and A. Grover,
“Climax: A foundation model for weather and climate,” arXiv preprint
arXiv:2301.10343, 2023. Martha Arbayani Zaidan (Senior Member, IEEE)
[209] J. Pantelic, Y. J. Son, B. Staven, and Q. Liu, “Cooking emission control received the Ph.D. degree in Automatic Control
with IoT sensors and connected air quality interventions for smart and and Systems Engineering from the University of
healthy homes: Evaluation of effectiveness and energy consumption,” Sheffield. He was a Postdoctoral Research Associate
Energy and Buildings, vol. 286, p. 112932, 2023. at the University of Maryland, a Fellow at Aalto
[210] R. He, R. L. Tiong, Y. Yuan, and L. Zhang, “Enhancing resilience University, and a Research Associate Professor at
of urban underground space under floods: Current status and future Nanjing University. Currently, he is an Academy Re-
directions,” Tunnelling and Underground Space Technology, vol. 147, search Fellow and Data Scientist at the Department
p. 105674, 2024. of Computer Science and Institute for Atmospheric
[211] Z. Lv and W. Shang, “Impacts of intelligent transportation systems on & Earth System Research (INAR), University of
energy conservation and emission reduction of transport systems: A Helsinki. His research interests include artificial
comprehensive review,” Green Technologies and Sustainability, vol. 1, intelligence and machine learning for intelligent control systems, health
no. 1, p. 100002, 2023. monitoring technologies, applied physics, and atmospheric and environmental
[212] D. Zhuang, V. J. Gan, Z. D. Tekler, A. Chong, S. Tian, and X. Shi, sciences.
“Data-driven predictive control for smart hvac system in iot-integrated
buildings with time-series forecasting and reinforcement learning,”
Applied Energy, vol. 338, p. 120936, 2023.
 27

Naser Hossein Motlagh received the D.Sc. de- Tuukka Petäjä is a Professor of Experimental At-
gree in Networking Technology from the School of mospheric Sciences and Vice-Director of INAR at
Electrical Engineering, Aalto University, Finland, in the University of Helsinki. He earned his Ph.D.
2018. He is an Academy of Finland Research Fellow (2006) and Docent (2011) in Physics from the
at the Department of Computer Science, University University of Helsinki and completed postdoctoral
of Helsinki and at the Nokia Center for Advanced research at the U.S. National Center for Atmospheric
Research (NCAR). Previously, he was a Postdoctoral Research (NCAR). He oversees INAR’s aerosol do-
Fellow within the Helsinki Institute for Informa- main and research infrastructures. He has over 430
tion Technology HIIT and Helsinki Center for Data peer-reviewed articles, including 10 in Science and
Science (HiDATA). His research interests include 9 in Nature, with over 22,000 citations and an H-
the Internet of Things, wireless sensor networks, factor of 69. Recognized as a highly cited scientist
environmental sensing, smart buildings, and unmanned aerial and underwater since 2014, his awards include the FAAR Award, Vaisala Award (2013), and
vehicles. Science and Technology in Society Future Leader Award. He also leads the
BAECC campaign, directs PEEX, serves on the PACES initiative board, and
is an academician in the International Eurasian Academy of Sciences.

Petteri Nurmi is a Professor of Computer Science

at the University of Helsinki, where he leads the
Pervasive Data Science Group. He holds a Ph.D.
(2009) and M.Sc. (2006) in Computer Science from
the University of Helsinki, and was previously a
Lecturer at Lancaster University. His research fo-
cuses on pervasive computing, IoT, data science,
and environmental monitoring, with a strong em-
phasis on intelligent mobile systems, sustainable
computing, and novel sensing applications. He has
published over 150 articles, including several in
flagship conferences, and is highly active in international collaborations,
journal publications, and conference committees.

Tareq Hussein is a Professor at the Institute for

Atmospheric and Earth System Research (INAR), Sasu Tarkoma (Senior Member, IEEE) is a Profes-
University of Helsinki, and a Professor at the Uni- sor of Computer Science and Dean of the Faculty of
versity of Jordan. He holds a Ph.D. in Physics (2005) Science at the University of Helsinki. His research
and is a Docent in Physics from the University focuses on edge intelligence, IoT, and distributed
of Helsinki. His research focuses on atmospheric systems. He has 472 publications, 14,463 citations,
and environmental sciences, with a strong emphasis and an H-index of 57 (AD Scientific Index 2025).
on air pollution, urban and indoor air quality. This Author of 4 textbooks, he holds 11 US patents and
includes the dynamics and physical characterization over 20 international patent applications. Professor
of aerosol particles, emissions, dry deposition, and Tarkoma has received numerous Best Paper awards
health effects. (IEEE PerCom, ACM CCR, ACM OSR) and chairs
the Finnish Scientific Advisory Board for Defence
(MATINE). He is a Fellow of IET and EAI.

Markku Kulmala who leads INAR, is a highly

influential researcher in atmospheric science. He’s
known for his work on global measurement net-
works and air quality-climate change interactions,
particularly through SMEAR and GlobalSMEAR.
With over 1100 original research papers, including
18 in Nature and 17 in Science, he has over 60,000
citations and an H-factor of 120. Formerly ranked
first in Geosciences citations, he is a foreign member
of the Chinese and Russian Academies of Sciences,
a member of five other academies, and president
of the European Center of International Eurasian Academy of Sciences.
He has received over 10 international awards and holds nine honorary
doctorates/professorships.