This article has been accepted for inclusion in a future issue of this magazine.

Content is final as presented, with the exception of pagination.

NEXT GENERATION INTERNET

Toward Sustainable 6G leveraging Digital

Twin and Artificial Intelligence:
Framework and Case Study
Dianlei Xu, Xiang Su, Sasu Tarkoma, and Pan Hui

Abstract tions but demand substantial power for sig-

The forthcoming 6G networks will adopt nal processing
higher frequency bands, ultra-dense base sta- • Native edge intelligence (EI) support, where
tion deployments, and advanced technologies, edge servers (ESs) provide local computation-
such as massive MIMO and edge intelligence, al resources for network optimization, result-
to achieve enhanced bandwidth, connectivity, ing in significant energy consumption [2].
and intelligence. However, these advancements Sustainability is a key objective in the devel-
will lead to significant increases in energy con- opment of 6G networks, with energy efficiency
sumption. In this article, we propose to integrate and carbon footprint reduction serving as critical
digital twin and artificial intelligence to systemat- priorities [3]. However, the substantially increased
ically solve the problem of energy sustainability power requirements of 6G networks present a
in 6G networks. We explore the key challenges significant challenge to operators’ sustainability
of integrating these two technologies and design goals. Consequently, operators are intensifying
an intelligent digital twin framework to address efforts to balance these high energy demands
them. Our framework adopts a modular design, with their commitment to sustainable develop-
encompassing modeling, simulation, optimization, ment targets [4].
and long-term evolution modules. As a case study, A promising solution to address these challeng-
we conduct a preliminary implementation of the es lies in the integration of Digital Twin (DT) and
framework in the 5th Ring Road area of Beijing, Artificial Intelligence (AI), as their complementary
China, demonstrating its effectiveness in improv- relationship enhances each other’s capabilities.
ing energy efficiency. Experimental results show DT is the digital replica of the physical world,
that our framework significantly reduces the ener- while AI encompasses intelligent algorithms for
gy consumption of 6G networks while optimizing learning and decision-making from big data. On
resource allocation to reduce computing latency. the one hand, AI helps DTs learn complex behav-
ior patterns of physical entities for accurate envi-
Introduction ronmental modeling. On the other hand, DTs can
The deployment of fifth-generation (5G) networks simulate various real-world scenarios with negli-
enables various applications like telemedicine, gible cost using generative algorithms, providing
smart factories, and autonomous vehicles through rich datasets for AI training. The DT-AI integra-
its high-speed, low-latency, and massive connec- tion enables continuous learning and optimiza-
tivity capabilities. Building on these achievements, tion of 6G networks for energy efficiency through
both industry and academia are now focusing on real-time monitoring and decision-making. By
sixth-generation (6G) networks to further enhance leveraging AI algorithms on DT-provided data-
performance and intelligence. sets, operators can proactively optimize resource
Compared with 5G networks, which consume allocation and implement dynamic power-saving
three to four times more energy than 4G, 6G net- strategies. This integration allows the simulation
works are expected to have substantially higher of energy-efficient configurations prior to physi-
energy demands [1]. This stems from: cal implementation, thereby reducing real-world
• High-frequency bands (mmWave, THz, and trial costs. Consequently, the DT-AI integration
VLC), which require more power due to enhances operators’ ability to manage urban net-
greater propagation losses despite enabling work complexities while improving energy sus-
higher data rates tainability in 6G systems through intelligent and
• Ultra-dense base station (BS) deployments adaptive network management.
needed for coverage and capacity, leading Existing DT frameworks [5, 6] for 6G networks
to increased overall energy consumption face several limitations in urban scenarios:
• Massive MIMO technology, where multiple • Their rigid mathematical and physical models
antennas support numerous device connec- inadequately capture complex urban dynamics.

ISSN: 0163-6804 The authors are with University of Helsinki, Finland; Pan Hui is also with The Hong Kong University of Science and Technology,
Digital Object Identifier: 10.1109/MCOM.003.2400389 Hong Kong. Xiang Su is the corresponding author.

1 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ IEEE Communications Magazine • Accepted for Publication
 Aspect Our framework Ref. [5] Ref. [6]

Predominantly physics- and mathematical

Modeling Data-driven, learning-based Mathematical and mechanism-driven
model-based

Simulation Generative AI-enhanced realistic simulation Simulations with predefined assumptions Basic simulation models for specific systems

Optimization Dynamic self-adaptive optimization Closed-loop adaptive optimization Fixed optimization rules, manual adjustments

Evolution Automatic long-term adaptation Not considered Not considered

Intelligence Full-process AI empowerment Partial use of AI in optimization Primarily relies on traditional methods

Focus 6G energy sustainability Factory network Planning and deployment 6G radio testing

TABLE 1. Comparison of different DT frameworks.

• Their simulation modules rely on simpli- patterns vary by application scenarios and net-
fied numerical assumptions, creating gaps work conditions — from video streaming APPs
between simulated and real-world scenarios demanding high bandwidth to autonomous vehi-
• They maintain static structures post-deploy- cles requiring extensive computational resources
ment, overlooking the long-term evolution for navigation and sensor processing.
needs of DTs as cities grow and communica-
tion patterns change. Simulating Real-World Scenarios
In this article, we address these limitations Simulating the results of various adjustments on
by designing an intelligent DT framework that physical 6G networks is crucial for ensuring DT
systematically integrates AI and DT to improve system robustness and reliability. However, such a
energy sustainability in 6G networks. Our frame- simulation faces the following challenges.
work fundamentally differs from existing solutions Diverse Realistic Scenarios: Real-world urban
through its data-driven, learning-based philosophy environments encompass diverse user behaviors
and full-process AI integration. Table 1 highlights and applications, from indoor office work to out-
the key differences between our designed frame- door activities. Conventional simulators often rely
work and existing ones. Our framework encom- on oversimplified assumptions and limited scenarios.
passes four synergistic modules that enable the For instance, random walk models in OMNeT++ [7]
DT to not just mirror the current network state fail to capture urban dynamics. The lack of environ-
but to truly understand and predict network mental variability in simulations leads to less realistic
dynamics: a modeling module that learns to cap- outcomes and inaccurate performance predictions.
ture complex network patterns from real-world Interplay Modeling: Urban environments exhib-
data, a simulation module that can generate and it complex interactions across domains, including
explore potential scenarios via generative AI, an resource competition among users and collaborative
agent module that implements self-adaptive opti- behaviors of autonomous vehicles. Conventional
mization based on learned patterns, and an evo- simulators struggle to integrate these cross-domain
lution module that enables continuous learning as interactions, resulting in isolated models. For exam-
the environment changes. Our case study results ple, the BonnMotion simulator [8] generates mobile
demonstrate that this integrated approach sig- node trajectories using simplified models like the
nificantly reduces both energy consumption and Manhattan grid, failing to capture real-world com-
computational latency in 6G networks, highlight- plexities such as traffic light influences.
ing its potential to enhance both sustainability and
performance in future network environments. Global Optimization
The integration of DT and AI aims to optimize
Key Challenges of Integrating DT and AI energy sustainability and user experiences in phys-
ical networks. However, achieving global opti-
Model Creation for the Physical Twin mization is challenging due to intricate urban
In this integration, modeling the behavior patterns environments, large-scale networks, diverse user
of entities, for example, mobile users, vehicles, demands, and heterogeneous mobility patterns.
and Internet of Thing (IoT) devices, within urban Large-Scale Optimization: Urban areas con-
environments presents significant challenges due tain numerous mobile users, vehicles, IoT devices,
to the following factors. and infrastructures. Optimization involves man-
Complex Urban Environments: Urban envi- aging interactions and cooperation among these
ronments are characterized by dense buildings entities - ESs collaborate for load balancing while
and skyscrapers, which significantly affect wireless users compete for resources. The multitude of
signal propagation through blockage, reflection, variables and constraints complicates real-time
and diffraction. The placement of infrastructure, optimization, which is crucial for adapting to
both indoors and outdoors, adds to this com- changing urban conditions.
plexity. The physical environments remain highly Heterogeneous User Demands: Mobile users,
dynamic due to user mobility, fluctuating traffic vehicles, and IoT devices have diverse service
loads, and varying environmental factors. demands. Mobile users require high bandwidth
Heterogeneous Behavior Patterns of Physi- for video streaming, vehicles need ultra-low laten-
cal Entities: Mobile users, IoT devices, and vehi- cy for autopilot, and IoT devices demand efficient
cles exhibit complex mobility patterns, particularly data transmission. These demands vary with time
as users transition between indoor and outdoor and location, requiring highly adaptive optimiza-
environments. Communication and computation tion algorithms across different scenarios.

IEEE Communications Magazine • Accepted for Publication This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 2
 Digital twin
Agent
Network slicing optimization Task offloading
Reinforcement Learning

Communication

Computation
Traffic load balancing Computation load balancing
Unsupervised Learning
Spectrum management Resource allocation
Edge Caching Solve Supervised Learning Solve Light-weight AI models
BS switching strategy Traditional Methods ES switching strategy

Update Feature Feedback Trial

Evolution Model Simulation

Synchronization
GAN

Fine-tune VAE

Update Transformer
Integration Learn Feature Generate
Diffusion
Expansion

Data Action

Physical twin
Data collection

FIGURE 1. Illustration of intelligent DT framework. The model module transforms data into digital representations of physical entities. The simulation module
uses generative AI for cost-free scenario simulation. The agent module applies algorithms to optimize 6G energy sustainability. The evolution module
ensures the DT adapts to physical world changes, continuously improving performance.

Diverse and Dynamic Mobility Patterns: integrate comprehensive datasets. Scaling these
Mobile users, vehicles, and IoT devices exhibit capabilities is crucial for managing growing urban
diverse mobility patterns [9], affecting connectiv- system complexity.
ity, network load, and resource allocation. Stat-
ic optimization strategies prove insufficient. For Framework Design
example, a fast-moving vehicle may fail to off- To address these challenges systematically, we
load tasks to the nearest ES, leading to repeated design an intelligent DT framework for sustain-
requests and resource waste. Thus, optimization able 6G networks, as shown in Fig. 1. The sys-
algorithms must account for mobility dynamics to tem comprises four main modules in DT: model,
enhance service quality and energy efficiency. simulation, agent, and evolution, which operate
in a highly coordinated manner through multiple
The Evolution of DT feedback loops. The model module transforms
The DT must continuously learn from incoming physical-world data into digital knowledge, which
data, adapt to changes, and evolve over time to the simulation module uses to generate and val-
maintain its accuracy and effectiveness. idate scenarios. The agent module then makes
Model Adaptation: As urban physical systems data-driven decisions based on these insights,
evolve with changing user patterns and external while the evolution module ensures synchroni-
conditions, DT models may experience accura- zation between digital and physical twins. These
cy degradation. Urban environments change over modules maintain continuous interaction through
time due to technological advancements and life- well-defined feedback mechanisms.
style shifts, affecting service demands and resource
allocation. For example, the rise of remote work Model Module
alters commuting patterns, impacting service The model module in the DT framework trans-
requirements and infrastructure deployment. forms raw data from the physical world into action-
Urban environments also evolve through new able knowledge. By leveraging trajectory data from
developments and infrastructure projects, affecting mobile users, vehicles, and IoT devices, it learns
traffic flow and signal propagation. Static DT mod- and predicts mobility patterns. Smartphone data
els struggle to capture these evolving dynamics. reveals physical activities, while network usage
Scalability: The growth in population, vehicles, and app interaction data inform models of mobile
and IoT devices leads to increased data volume network behaviors, enhancing user experiences in
and computational demands, complicating global communication and computing.
optimization. As physical systems become more The module adopts learning-based methods
sophisticated, DTs must incorporate detailed for modeling to effectively capture the heteroge-
models, support high-fidelity simulations, and neous behavior patterns of various network enti-

3 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ IEEE Communications Magazine • Accepted for Publication
ties. Spatial-temporal features in trajectory data framework as shown in Fig. 1. It receives learned
can be extracted using convolutional neural net- patterns and features from the model module as This comprehensive mod-
works (CNNs), while transformer networks cap- the foundation for scenario generation, ensuring eling approach effectively
ture long-term dependencies to create mobility simulated environments accurately reflect real-
predictions. Network usage patterns can be mod- world dynamics. The module maintains a bidi- addresses the challenge of
eled through time series analysis, with recurrent rectional relationship with the agent module: it complex urban environments
neural networks (RNNs) forecasting demand. provides a safe testing environment for agent with dense infrastructure and
For physical environments, the module inte- strategies while receiving trial actions for valida-
grates 3D modeling techniques using high-res- tion, offering detailed performance analytics when dynamic factors.
olution map data, point cloud processing, and sub-optimal results are detected. This enables
mesh generation. Real-time data streams, such as rapid algorithm refinement before physical
traffic flow and weather information, dynamically deployment. As the physical environment evolves,
update these models. In 6G networks, ray-tracing the simulation module receives updated features
is adopted to simulate wireless signal propaga- through the model module, which is synchronized
tion, considering building materials, weather, and with physical changes via the evolution module.
obstacles, enabling precise modeling of urban sig- Simulation outcomes are continuously compared
nal environments. This comprehensive modeling with actual results when available, creating a com-
approach effectively addresses the challenge of prehensive feedback loop that improves simula-
complex urban environments with dense infra- tion accuracy while reducing operational risks.
structure and dynamic factors.
The module maintains model accuracy through Agent Module
continuous validation against real-world measure- The agent module functions as the self-adap-
ments. Performance metrics are monitored in real tive decision-making core of the DT framework,
time, triggering model updates when accuracy dynamically adjusting optimization strategies based
degrades. Beyond internal validation, the model on network states and performance feedback. It
module actively interacts with other components combines different algorithms (reinforcement learn-
of the framework, as shown in Fig. 1. It feeds ing, supervised learning, unsupervised learning, and
learned features to the simulation module for sce- traditional methods) to leverage complementary
nario generation and provides predictive insights strengths. This approach allows us to decompose
to the agent module for decision optimization. complex optimization problems into manageable
The effectiveness of these predictions is constant- sub-problems for efficient solutions. For instance,
ly validated through feedback from the agent’s unsupervised learning partitions the network into
actions and their outcomes. Meanwhile, the evo- sub-regions, while reinforcement learning opti-
lution module ensures the models remain syn- mizes resource allocation within each region. The
chronized with physical-world changes through its module dynamically evaluates network conditions
fine-tuning mechanism. and adjusts optimization objectives in real-time,
integrating state information with predictive insights
Simulation Module from the model module. This mechanism effective-
The simulation module is to generate diverse ly accommodates heterogeneous user demands
real-world scenarios and validate agent deci- and diverse mobility patterns.
sions before deployment. The simulation process The agent module orchestrates complex inter-
begins by receiving learned patterns from the actions with other components, as shown in Fig.
model module, which form the basis for scenario 1. It uses predictive insights from the model mod-
generation. These patterns include user mobil- ule to inform decision-making. A bidirectional
ity trajectories, network usage behaviors, and “trial-feedback” loop with the simulation module
environmental dynamics. The module employs validates strategies in simulated environments,
different generative AI techniques based on spe- refining them based on performance analytics
cific simulation requirements; for example, GANs before deployment. The agent’s actions are syn-
excel at creating realistic urban mobility scenarios, chronized with the evolution module to ensure
VAEs efficiently generate network load patterns, physical twin modifications are mirrored in the
Transformers model temporal dependencies in DT. Algorithms undergo rigorous training and vali-
user behaviors, and Diffusion models synthesize dation but remain adaptable through online learn-
complex environmental conditions. This diverse ing mechanisms. This interaction system enables
set of generative AI techniques enables the sim- autonomous energy sustainability optimization
ulation of various realistic scenarios that conven- across varying network conditions and scales,
tional simulators struggle to capture. ensuring strategy adjustments are made when
The simulation execution follows a systematic performance deviates from expectations or when
approach to handle the challenge of modeling network conditions change significantly.
complex interactions. The module first creates
base scenarios incorporating known patterns and Evolution Module
constraints. Then, it introduces variations through The evolution module synchronizes the DT with
controlled parameter adjustments, simulating dif- the evolving physical twin through four key mech-
ferent operational conditions such as traffic load anisms: synchronization, fine-tuning, integration,
fluctuations, infrastructure changes, and unex- and expansion. When the agent makes deci-
pected events. The simulation considers multiple sions and applies actions on the physical twin,
timescales, from immediate responses to long- modifications (e.g., infrastructure deployment
term evolution, enabling comprehensive evalua- or operational adjustments) must be mirrored in
tion of energy-saving strategies. the DT. External events, such as weather anom-
The module implements comprehensive inter- alies or geological disturbances, can significantly
action mechanisms with other components of the alter physical environments. The synchronization

IEEE Communications Magazine • Accepted for Publication This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 4
 data collected from one of the major Chinese
Closed Edge Server Closed Computing Flow
BS3 ISPs during April 20–26, 2016, including mobility
Edge Server Computing Flow patterns and traffic demand of 286,665 users as
well as operational data from 3,880 BSs, and then
Closed BS Closed Traffic Flow
used to model the mobility and traffic demand
BS Traffic Flow of mobile users; and Agent Module, where a
ES3
BS2
deep reinforcement learning (DRL)-based agent
u3 u4 is designed to minimize 6G network energy con-
BS1
sumption and computing latency through coordi-
nated control of BSs and ESs.
u6
u1
u2 ES2 Problem Description
ES1 u5 The scenario is shown in Fig. 2. In an urban envi-
ronment, each BS is equipped with an ES, pro-
viding users with communication services and
FIGURE 2. Illustration of cooperation among (BS, ES) pairs. Ultra-dense urban deployment results in lots of BS overlap. computational resources. Users engage with
ESs on BSs jointly provide traffic and computing services. BS3 and ES3 are closed with u3 and u4 switched to (BS1, mobile applications like TikTok, consuming both
ES1) and BS2 respectively. ES2 is closed due to low computational demand. communication and computational resources.
Tasks are offloaded from mobile devices to the
mechanism addresses these changes through nearest ES via the connected BS. Users can only
continuous environmental monitoring via sensor utilize services from the ES associated with their
networks and real-time updates to the digital rep- connected BS. The ultra-dense deployment of BSs
resentation, which are propagated to other mod- creates overlapping coverage areas, offering users
ules and validated with performance indicators. multiple connectivity options. In order to reduce
As the physical environment evolves, DT energy consumption, underutilized BSs and ESs
models must adapt. The fine-tuning mechanism can be shut down, with users redirected to neigh-
employs adaptive learning approaches, including boring BSs and ESs that have sufficient capacity.
incremental learning to integrate new patterns For instance, BS3 and ES3 could be closed, with
while retaining existing knowledge, transfer learn- their neighbors serving mobile users u3 and u4,
ing for efficient adaptation, and continuous valida- as shown in Fig. 2. The objective is to minimize
tion to maintain performance. This ensures models overall energy consumption while maintaining
and algorithms in the model, agent, and simulation service quality through (BS, ES) pair cooperation.
modules adapt to environmental changes. This optimization involves both BS/ES closure
Although the DT does not exhaustively rep- decisions and effective user load reallocation to
resent the physical world, it is designed for con- maintain connectivity and computing support.
tinuous expansion. The integration mechanism
supports growth by establishing standardized Agent Design
interfaces for external DT connections, har- To address the problem, we propose a reinforce-
monizing data formats, and coordinating deci- ment learning-based algorithm for optimal (BS,
sion-making across multiple DTs. The expansion ES) pair switching. We model this as a Markov
mechanism enables the dynamic addition of new Decision Process, where the agent makes deci-
features, incorporation of new data sources, and sions, receives rewards, and transitions between
scalable resource allocation strategies. states, as shown in the lower part of Fig. 3. Similar
These four mechanisms collaborate with other to chess gameplay, where players sequentially
framework components to address model adap- place pieces, our algorithm processes a random
tation and scalability challenges, as shown in access sequence of (BS, ES) pairs. At each step,
Fig. 1. When environmental changes occur, the the agent decides an action for a specific pair
synchronization mechanism updates the digital based on observed conditions and receives a
representation and informs the model module. calculated reward. This process continues until
These updates propagate through the framework, the last pair, enabling systematic evaluation of
shaping simulation scenarios and agent decisions. shutdown impacts while optimizing energy con-
The fine-tuning mechanism refines algorithms and sumption and load reallocation. Our algorithm
parameters, while the integration and expansion builds on the PPO architecture [11], comprising
mechanisms facilitate system growth and external an actor-network for action generation and a crit-
DT connections. This interaction ensures synchro- ic network for performance evaluation. As shown
nization with physical world changes while main- in the upper part of Fig. 3, both networks utilize
taining scalability and performance. two-layer graph convolutional networks [11] and
two-layer fully-connected networks for feature
Case Study extraction, followed by a two-layer fully-connect-
In this section, we demonstrate the potential of ed network for output. The state space, action
our intelligent DT framework through a case study space, and reward function are defined as follows.
conducted in the 5th Ring Road area of Beijing, State Space: The state comprises four com-
China. The current implementation of our frame- ponents (Requests, Neighbors, Neighborstatus,
work focuses on three key modules to assess Masks), as shown in Fig. 3. Requests represent
its feasibility and effectiveness: Model Module, mobile users’ traffic and computation requests for
where a 3D model of the physical environment is the current (BS, ES). Neighbors denotes the neigh-
constructed using building data from OpenStreet- bor set of the current (BS, ES). Neighborstatus
Map, and ray tracing is applied to simulate wire- indicates neighbors’ active/closed status. Masks
less signal transmission; Simulation Module, where represents the index of (BS, ES) and the current
the simulator in [10] is fine-tuned using real-world sequence step.

5 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ IEEE Communications Magazine • Accepted for Publication
 Action Space: Three possible actions exist for
each (BS, ES): (active, active), (active, closed), and Values Actions
(closed, closed).
Reward Function: The reward, calculated only
at the sequence’s last step, is a weighted sum of Linear Layers
overall energy consumption and average comput-
ing latency combined using two hyperparameters. GCN Layers GCN Layers Linear Layers Linear Layers
All rewards before the last step are zero.
We compare our algorithm with two baselines:
the original strategy without optimization and the (2,3)
Efficiency-First Strategy (EFS) from [12], which
greedily selects BSs with the highest energy effi- Requests Neighbors Neighbor status Masks
ciency in each iteration. Obs. Obs. Act. Obs. Act. Obs. Act. Obs. Act.
Act.
Experiments and Results
To categorize network conditions based on traffic
load, we define seven levels of network states,
corresponding to average traffic demands per
—
time slot for users, denoted as T. The length of
a time slot is 10 seconds. Based on our empiri-
cal evaluation, the agent can infer for all (BS, ES) FIGURE 3. The illustration of DRL-based (BS, ES) switching algorithm. The lower part shows a randomly generated
pairs in approximately 1.3 seconds, enabling fre- visiting sequence. The agent decides the status of each (BS, ES) pair by this sequence. The upper part illustrates
quent status updates. The update frequency can the structure of the DRL algorithm.
be flexibly adjusted according to network dynam-
ics and operational requirements, with our imple-
mentation adopting a 10-second update interval
to balance responsiveness and system stability.
We assume the ESs adopt the same model as the
Dell R730. We adopt the same power model of
R730 as [13]. Considering that task arrival rates at
computing servers typically follow a Poisson dis-
tribution with an average rate of l, and that the
computing load of the tasks adheres to an expo-
nential distribution [14], we generate computing
tasks for mobile users with these two distributions
for each time slot. The average computing load of
tasks is set at 30 MB. Similarly, we set seven levels
to reflect different intensities of computational
task loads. We separately run the agent for one
hour under different network conditions.
The line plot in Fig. 4 presents the energy con-
sumption under various network conditions.1 It is FIGURE 4. Comparison of energy consumption (total energy used by all BSs and ESs in one hour) and ES resource
evident that our DRL-based algorithms achieve utilization (average CPU usage during task execution) across different network conditions by different solutions.
significantly lower energy consumption than the The line plot shows energy consumption, while the bar chart depicts ES resource utilization for each algorithm
other two baselines. Specifically, our approach under the same conditions.
can reduce energy consumption by up to 73 per-
cent compared to the non-optimized strategy.
The bar chart in Fig. 4 shows the performance of the computing latency as a case study. Results
resource utilization. Our DRL approach always indicate significant reductions in overall ener-
achieves the highest resource utilization. Users gy consumption and average computing time,
originally connected to closed BSs or ESs are real- as well as improvements in resource utilization,
located to other operational units. Consequently, demonstrating the potential of our framework to
our strategy not only improves resource utilization enhance both sustainability and performance in
but also achieves substantial energy savings and future 6G network environments.
lower response latency, as shown in Fig. 5. This Currently, the preliminary framework is imple-
demonstrates the efficiency and robustness of our mented in a centralized manner on cloud serv-
designed framework in the context of dynamic ers. The DT system’s data storage and model
urban 6G networks. training processes are fully conducted in the
cloud, leveraging the cloud’s extensive compu-
Conclusion and Discussion tational and storage resources. Several practical
In this article, we discuss the key challenges of challenges need consideration for real-world
integrating DT and AI to solve the energy sus- deployment. The urban-scale DT system faces sig-
tainability problem in 6G networks and design an nificant hardware constraints, including massive
intelligent DT framework to systematically address storage requirements for data generated from 1 The calculation of energy
these challenges. The framework adopts a mod- user mobility, network states, and environmental consumption, resource utili-
ular design with essential components to ensure monitoring. AI empowerment demands significant zation and computing latency
robust and adaptive network management. To computational and energy resources, particularly are omitted due to space limit.
validate our proposed framework, we construct a in simulation processes. The energy footprint of Readers interested in detailed
methodologies can refer to
DT and design a DRL-based agent to jointly opti- maintaining the DT system, including model train- [15] for comprehensive infor-
mize the switching policy of 6G BSs and ESs, and ing, continuous data processing, and real-time mation.

IEEE Communications Magazine • Accepted for Publication This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ 6
 Cluster Computing, vol. 26, no. 3, 2023, pp. 1845–75.
[3] M. Matinmikko-Blue, S. Yrjölä, and P. Ahokangas, “Multi-Per-
spective Approach for Developing Sustainable 6G Mobile
Communications,” Telecommun. Policy, vol. 48, no. 2, 2024,
p. 102640.
[4] Y. Ding et al., “Carbon Emissions and Mitigation Potentials
of 5G Base Station in China,” Resources, Conservation and
Recycling, vol. 182, 2022, p. 106339.
[5] D. Xia et al., “Digital Twin and Artificial Intelligence for Intel-
ligent Planning and Energy-Efficient Deployment of 6G Net-
works in Smart Factories,” IEEE Wireless Commun., vol. 30,
no. 3, 2023, pp. 171–79.
[6] H. Gao et al., “Digital Twin Enabled 6G Radio Testing: Con-
cepts, Challenges and Solutions,” IEEE Commun. Mag., vol.
61, no. 11, 2023, pp. 88–94.
[7] A. Varga, “Omnet++,” Modeling and Tools for Network Simu-
lation, Springer, 2010, pp. 35–59.
[8] N. Aschenbruck et al., “Bonnmotion: A Mobility Scenario
Generation and Analysis Tool,” Proc. 3rd ICST, 2010.
[9] D. Xu et al., “IoT vs. Human: A Comparison of Mobility,” IEEE
TMC, vol. 21, no. 4, 2022, pp. 1257–73.
FIGURE 5. Comparison of computing latency (average delay from task reception to completion on ESs) across differ- [10] J. Gong et al., “Scalable Digital Twin System for Mobile
ent network conditions by different solutions.} Networks With Generative AI,” ACM MobiSys, 2023, pp.
610–11.
[11] J. Schulman et al., “Proximal Policy Optimization Algo-
simulation, poses additional challenges to ener- rithms,” arXiv preprint arXiv:1707.06347, 2017.
[12] N. Yu et al., “Minimizing Energy Cost by Dynamic Switch-
gy sustainability. These challenges become more ing on/Off Base Stations in Cellular Networks,” IEEE TWC,
pronounced as the network scales up, making it vol. 15, no. 11, 2016, pp. 7457–69.
difficult to maintain model accuracy and system [13] W. Wu et al., “A Power Consumption Model for Cloud
responsiveness while managing energy efficiency. Servers Based on Elman Neural Network,” IEEE TCC, vol. 9,
no. 4, 2019, pp. 1268–77.
A potential solution to these issues is a [14] K. Li, “Improving Multicore Server Performance and Reduc-
cloud-edge continuum architecture. Under this ing Energy Consumption by Workload Dependent Dynamic
approach, data processing and storage are hierar- Power Management,” IEEE TCC, vol. 4, no. 2, 2015, pp.
chically distributed: edge nodes handle real-time 122–37.
[15] B. Mao et al., “AI Models for Green Communications
data processing and temporary storage, while the Towards 6G,” IEEE Commun. Surveys & Tutorials, vol. 24, no.
cloud manages long-term data archival and com- 1, 2021, pp. 210–47.
plex analytics. Computational load and energy
consumption are optimized by performing tasks Biographies
like model training and large-scale simulations in Dianlei Xu is currently a joint doctoral student in the Depart-
energy-efficient cloud data centers, while laten- ment of Computer Science, Helsinki, Finland and Department
of Electronic Engineering, Tsinghua University, Beijing, China.
cy-sensitive operations remain at the edge. Such His research interests include edge computing, edge intelligence
a hybrid approach balances computational load, and reinforcement learning.
storage needs, and energy consumption while
maintaining system responsiveness. However, Xiang Su [M] is an associate professor at the Department of
Computer Science and Department of Agricultural Sciences,
careful investigation is required to manage trade- University of Helsinki, Finland. He was an Associate Professor at
offs between system performance, resource allo- Norwegian University of Science and Technology, Norway. He
cation, response latency, and energy efficiency. has extensive expertise on edge intelligence, extended reality,
and IoT.
Acknowledgments Sasu Tarkoma [SM] is a Professor of Computer Science and
This work was partly supported by the Research Dean of the Faculty of Science at the University of Helsinki,
Council of Finland University profiling funding 7 Finland. He is affiliated with the Helsinki Institute for Informa-
InterEarth RESET, NordForsk through the funding tion Technology (HIIT) and the Finnish Center for AI (FCAI).
He is the chairman of the Finnish Scientific Advisory Board for
to Nordic University Cooperation on Edge Intelli- Defence (MATINE). His research interests include Internet and
gence, project number 168043. distributed systems, IoT, mobile computing, and AI.

Refer ences Pan Hui [F] is a Chair Professor of Computational Media and Arts
[1] F. Salahdine, T. Han, and N. Zhang, “5G, 6G, and Beyond: thrust at the Hong Kong University of Science and Technology
Recent Advances and Future Challenges,” Annals of Tele- (Guangzhou), and Emerging Interdisciplinary Areas and Director
commun., vol. 78, no. 9, 2023, pp. 525–49. of the HKUST-DT Systems and Media Lab, Hong Kong University
[2] A. Katal, S. Dahiya, and T. Choudhury, “Energy Efficiency in Cloud of Science and Technology. He is an IEEE Fellow, an ACM Distin-
Computing Data Centers: A Survey on Software Technologies,” guished Scientist, and a member of Academia Europaea.

7 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ IEEE Communications Magazine • Accepted for Publication