See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/361718091

Knowledge graph construction for product designs from large CAD model
repositories

Article in Advanced Engineering Informatics · August 2022

DOI: 10.1016/j.aei.2022.101680

CITATIONS READS
49 926

2 authors:

Akshay Bharadwaj Binil Starly

North Carolina State University North Carolina State University
8 PUBLICATIONS 115 CITATIONS 130 PUBLICATIONS 5,671 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Akshay Bharadwaj on 07 November 2022.

The user has requested enhancement of the downloaded file.

 Knowledge Graph Construction for Product Designs from Large CAD Model Repositories
Akshay G. Bharadwaj, Binil Starly*
Edward P. Fitts Department of Industrial & Systems Engineering
North Carolina State University, Raleigh, NC 27695
Contact author: [email protected]

Abstract
Product Design based Knowledge graphs (KG) aid the representation of product assemblies through
heterogeneous relationships that link entities obtained from multiple structured and unstructured sources.
This study describes an approach to constructing a multi-relational and multi-hierarchical knowledge
graph that extracts information contained within the 3D product model data to construct
Assembly-Subassembly-Part and Shape Similarity relationships. This approach builds on a combination
of utilizing 3D model meta-data and structuring the graph using the Assembly-Part hierarchy alongside
3D Shape-based Clustering. To demonstrate our approach, from a dataset consisting of 110,770 CAD
models, 92,715 models were organized into 7,651 groups of varying sizes containing highly similar
shapes, demonstrating the varied nature of design repositories, but inevitably also containing a significant
number of repetitive and unique designs. Using the Product Design Knowledge Graph, we demonstrate
the effectiveness of 3D shape retrieval using Approximate Nearest Neighbor search. Finally, we illustrate
the use of the KG for Design Reuse of co-occurring components, Rule-Based Inference for Assembly
Similarity and Collaborative Filtering for Multi-Modal Search of manufacturing process conditions.
Future work aims to expand the KG to include downstream data within product manufacturing and
towards improved reasoning methods to provide actionable suggestions for design bot assistants and
manufacturing automation.
Keywords: Digital Manufacturing, 3D Shape Search, 3D Recommendation, Deep Learning.
-------------------------------------------------------------------------------------------------------------------------------
1. Introduction
Model-Based Enterprise (MBE) and its specific form, the Model-Based Definition (MBD), has produced
an engineering workflow centered on 3D Computer-Aided Design (CAD) representation. In MBD, the
CAD model serves as the data source for all activities in the product lifecycle, with data from multiple
workflow processes contained within Product Lifecycle Management (PLM) systems. The advantage of
data-associativity in these systems have been used to integrate traditional product manufacturing with
newer technologies such as augmented reality [1], automation and robotics, and machine-readable
multi-physics simulations [2] powered by Artificial Intelligence (AI). Vast improvements in computing
and AI algorithms have been vital for organizations looking to streamline and integrate their design,
manufacturing and supply chain systems [3]. This transformative potential of AI to improve design has
long been recognized [4] [5], with robust knowledge representations being crucial to leveraging this
ability [6]. Recent industrial applications have shown the enormous advantages of integrating simulation,
manufacturing and robotic assembly workflows to obtain optimal design solutions through AI. The vast
corpus of existing design data in organizations have the potential to be utilized to change design from an
iterative process towards providing pre-optimized design suggestions.
Product Design involves optimization of multiple factors: performance, ease of manufacturing and
assembly, cost and aesthetics. The parameters controlling these aspects of design are application-specific,
either well defined or implicit, and requires domain expertise to apply towards good design practices. For
example, Design for Additive Manufacturing requires concurrent evaluation of geometry, material, and
mechanical properties during the design stage [7]. Present day commercial software has the capacity for
optimized design suggestions through constrained design space exploration. They combine existing
standard data representations such as STEP and statistical techniques such as design of experiments to
provide optimal design parameters. However, these methods have scope for improvement, such as
through automatic mining of pre-existing design data to provide more versatile recommendations.
Considering the large amount of data generated by enterprises and those available in the open domain, it
is important to create this linked data from existing semi-structured sources automatically as demonstrated
recently in [8], [9] and [10].
At the enterprise level, data workflows need to include multiple sub-systems including electronic and
software, each with their own ecosystems [11]. The tracking of product data thus spans multiple users,
systems and domains. Clearly, data-driven methods to access and utilize multi-disciplinary engineering
information requires organization and integration of data across multiple existing data-stores. Modern
product management workflows contain the necessary data and base level organization, but there is a
need for methods to exploit this valuable data for decision making and potential reuse of design
knowledge. The knowledge that is inherently present within the data-stores that contain vast amount of
information on product data can only be extracted if the product data models are linked with attributes
that define the model, and relationships that link any two digital 3D models.
Knowledge Graphs (KGs), which are graph representations of linked entities, can provide an efficient
representation of connections among multiple structured and unstructured data sources. KG-based
representations have been shown to mitigate issues in product development involving multi-disciplinary
knowledge extraction and recommendation [12]. However, a large majority of work in this field focuses
on textual KG representations. Generalized & open data sources for product design contain text
classifications, but product specific descriptions in the open domain are scarce and non-standardized. This
makes complex ontology-based methods that are commonly used to create KGs difficult to implement.
Secondly, text alone does not completely cover all possible descriptions of diverse 3D data seen in
designs, except in the case for standard components, which are readily available in existing databases.
Finding relationships between non-standard components, be it parts, sub-assemblies, or even entire
assemblies is more useful for design re-use in product development. One way to account for these issues
is to use numerical representations of product shape as the core of the KG and creating a flexible
representation of textual descriptions across the lifecycle built around these 3D model instances.
In this paper, we demonstrate a hybrid method for construction of Knowledge Graphs (KG) from large
Product Design repositories, implemented by leveraging existing product hierarchies and metadata from
publicly available 3D data sources. In addition, 3D Shape Similarities obtained through Neural Networks
are used in conjunction with Unsupervised Graph Clustering to aid the construction of search indices
within the design data repository, with the goal of linking data within the graph for efficient design search
and recommendation. The resulting Product Design Knowledge Graph and the data used for construction
is released as a Linked Open Knowledge base for product design and manufacturing [13], [14]. Finally,
use of the KG to provide efficient design recommendations from prior data is shown, along with its
potential to use KGs for complex decision making. The use of 3D CAD models as the focus of the KG
mirrors enterprise-level data representations; we believe that extensions of this framework with rich
multi-disciplinary data available at the organizational level can enable better engineering driven
multi-domain decision making.
2. Related Work on Knowledge Graphs and Data-Driven Design
2.1. Knowledge Graphs and its Industrial Applications
Well-known Knowledge Graphs (KGs) for real world facts derived from linked-open data on the web
have gained popularity for question-answering systems over the last decade, with notable examples being
YAGO [15], Wikidata [16] and Freebase [17]. They are collections of ‘triples’ of the form (entity,
relation, entity) represented by a directed graph structure, with nodes of multiple types representing
entities and edges representing multiple relation types between them. There has been a growth in the
development of specialized ontologies for KGs as part of a trend in the natural sciences for
domain-specific KG applications. The biomedical field has seen significant work in knowledge graph
construction techniques, with work by [18] and [19] being two recent examples of applying existing
textual domain data to create a knowledge graph framework. Materials science has also seen advanced
use of KG methods to connect material property relationships and relevant literature sources, connecting
implicit and explicit knowledge with decision making [20].
It is observed that the overwhelming majority of knowledge graphs are created using Natural Language
Processing (NLP) of existing textual data, for example in work by Zhou et al [21] and Myklebust et al
[22]. This can also be seen in the exhaustive review by Li et al [12], where neural network-NLP
approaches are used extensively to formalize knowledge for industrial products and services. This
approach may not be able to fully represent data in the product design domains, leaving out relations that
are not easily extracted in textual form. The large amount of multi-domain design data in enterprises
means that automated knowledge extraction and codification is crucial for further applications.

2.2. Knowledge Extraction from 3D Data using Neural Networks

Over the last decade, the application of Neural Networks (NNs) for classification and retrieval
applications has resulted in massive improvements in 3D shape representations. The use of image, point
cloud and voxel data to characterize 3D shapes has seen great success. 3D Shapenets [23], VoxNet [24]
and Pointnet [25] are some well-established techniques for 3D shape recognition using point cloud and
voxel data. The maturation of computer vision for image recognition has meant that these methods have
been successfully applied in the context of 3D shapes as well; the highly influential work in this field by
Su et al [26] that uses multiple views of 3D shapes is a notable example of the view-based 3D descriptors.
This was followed by work such as RotationNet [27] and our previous work on CAD model classification
using multi-view images [28]. Easy availability of image data means that NNs pre-trained on
well-established image data such as ImageNet [29] can be used to obtain accurate models trained on
relatively small datasets much quicker. This is a critical advantage when automated knowledge extraction
is applied at an industrial scale. In recent years, work on applying these techniques to CAD data for more
specific feature recognition (as opposed to representing the whole part) and shape simplification has
gained prominence as seen in work by Zhang et al [30] and Song et al [31]. These advantages of these
techniques for automated knowledge acquisition from 3D shapes are crucial to knowledge representation
of product designs. They provide a way to infer relations between currently unconnected data in the
design and manufacturing domains and provide more opportunities to improve data-driven design.
2.3. Data-Driven Design, Retrieval & Knowledge Graphs
In contemporary industrial applications, standardization of components has been widely used to enable
large-scale production while significantly reducing design effort. For example, in the aerospace industry,
there is an interest in reusing brackets from existing data for newer applications; in the work by Clark et al
[32], the use of hierarchical clustering based on bracket features and dimensions is demonstrated to find
representative parts from their entire dataset. Due to the restrictive nature of enterprise data, there has also
been great interest in knowledge structuring outside industry which is crucial to drive automated
knowledge acquisition for design [5]. Motivated by the same restrictions of enterprise data, there have
been multiple open design repositories of varying size and functional descriptions proposed and used in
[33]–[36] and the data used in this paper [37] to drive data science applications in product design.
The use of graphs and domain-specific ontologies for recording product data over the design and
production cycle has gained popularity due to its potential for multi-dimensional decision making [38]
[39] [40]. These include process modeling through MBE [41], automated assembly sequence generation
[42], design decision support [43], capturing intent and data reuse for physics-based simulation [44] [45].
These ontologies can be leveraged to create Knowledge Graphs (KGs), which have great potential for
performing search and inference on real-world data.
Text-based knowledge graphs, with data obtained from existing design documents have been a focus of
research over the last few years. The use of domain specific text-driven semantic knowledge graphs has
been explored for representing explicit and tacit design knowledge [46] and efficient process planning
using a Process Knowledge Graph obtained from CAD/CAM systems [47]. Helping designers retrieve
assemblies based on functional semantics was shown in the work by [48]. Multi-criteria topological
similarity was used by [49] to enable similarity assessment, decomposing assemblies into a graph with
common parts classified into node types and the kinematic assembly operation sequence denoting the
relation between them. An approach for applying graph-based knowledge reuse in product development
was proposed by [50], with the potential to vastly increase knowledge reuse in the manufacturing
industry. The problem of connecting multiple sub-systems containing associated data provides one major
barrier to the realization of this goal; [51] provides a solution to link and trace data throughout the product
lifecycle by creating digital threads across multiple system interfaces. Recently, a generalized solution to
track products and their manufacturing dependencies using graph databases was shown by Martinez-Gil et
al [52], which showed that product data in graph form outperformed relational databases. The advantages
of graph-based approach for the product domain are clearly seen from the literature.
In conjunction with textual data, several Knowledge Graph implementations in the product domain
include geometry as the basis for construction. These KGs use 3D data to inform further decision making
in design and manufacturing. As seen in the work by Ferrero et al [36], these approaches have the ability
to drive applications such as functional classification of components within assemblies, using data
obtained from CAD models. CAD model KG based on decomposition of 3D point cloud to primitive
shapes was applied for shape-based retrieval by [53], utilizing unsupervised methods to group 3D shape
primitives. KG driven assembly planning was demonstrated by Zhou et al [54], where information from
the CAD model and the assembly process document were used to generate assembly sequences. The
authors demonstrate that the use of 3D shapes is key to be able to account for potential collisions, correct
orientation, and associated manufacturing parameters. From the literature, systems for design intelligence
clearly need to combine multiple modalities: textual data, implicit/explicit knowledge, design intent,
simulation, manufacturing capabilities, and user requirements in order to enable true design intelligence.
An extensive search of existing KG literature in the design/manufacturing space by [55] shows that most
knowledge graphs are created using a top-down approach, involving the creation of an ontology which is
then populated with data to create the KG. However, an ontology-based approach is time consuming to
create and update; automating the acquisition, processing and use of knowledge has high value when
dealing with large amounts of diverse domain data [56]. Additionally, many manufacturing KGs are
modeled as Resource Description Framework (RDF) graphs, where nodes and edges do not contain
attributes. In contrast, the Labeled Property Graph (LPG) graph model can carry properties as key-value
pairs and has the potential to outperform RDF for complex graph queries [57], while creating a more
compact representation of properties. LPG and other ‘hyper-relational’ graph structures such as RDF*
show possibilities for efficient representation of multiple relationship details and weights. This form of
data representation is well-suited to exploratory approaches and sparse data.
Currently, many KGs in the product domain focus on one aspect of the product cycle such as design,
simulation, manufacturing or assembly. Many of these KGs are constructed using concept mapping &
classification of previous textual data as seen in [58]. However, the specialized nature of each of these
domain leads to siloed data sources with little or no interaction across the product lifecycle that inherently
will contain diverse textual descriptions. Work by Hao et al [59] answers this problem by the use of
generalizable and domain-independent Decision Support Problem Technique (DSPT) to generate multiple
KGs for domain-decision making processes, demonstrated on a supply-chain design domain. This method
provides a common knowledge template for designers, but interrelated sectors continue to be isolated
from another, which may lead to difficulty in multi-dimensional decision making.
Considering these limitations of text-based KGs, this paper describes a method using 3D shape & relevant
metadata as the center of the product description and decision-making process. This can pave the way for
proactive recommendations early in the product cycle, by using associations from prior data not explicitly
linked to design. Importantly, for current applications involving large amounts of data, inter-connection
between data must be created automatically and must be easily searchable. A method to achieve this is
shown using numerical representations or ‘embeddings’ of design similarity to create a graph.
Subsequently, unsupervised methods are applied to group parts into communities, and to account for the
large variety of products that do not fit under standard part categories. At the enterprise level, unified
product-based KGs thus created can be used for company-level data tracking, analysis and
recommendation from diverse sources using data-driven reasoning. Textual data associated with design
decisions using domain-specific ontologies can be linked to design documents, which can then generate
automated suggestions.
3. KG Construction Methodology
Knowledge Graph generation and exploitation is generally divided into Knowledge Extraction and
Construction, followed by Deduction of Patterns and Usage [12]. This section describes the methodology
of Data Extraction and KG Construction. Two possible approaches exist for KG construction: modeling
an ontology and then populating it with data (Top-Down), or by automatically inducing structure from the
data (Bottom-Up) [55][60]. Here, we propose a hybrid approach by extracting directly available textual
data for Top-Down relations, as well as the use of latent shape data representations for Bottom-Up
relationship linking.
3.1. Knowledge Extraction
Existing product design stores in the form of industry standard STEP files and their extensions are a
primary source of data and are used for day-to-day design decision making. Individual part design data is
represented in multiple formats with varying amounts of information, and the similarity between them can
be computed based on their associated text data. The existing technical text data and hierarchical
information from design files is the primary data extracted from the CAD files. However, these are not
always complete; obtaining comprehensive textual descriptions and associated data of parts is a laborious
process.
There has been a large body of work devoted to Neural Network Classification and Retrieval methods of
3D components based on shape ([23], [24]) as well as image-based methods, which have had success as
seen in [26], [27], [61]. These methods can generate feature vectors or embeddings as global
representations of the parts. Our method leverages these representations for further search and analysis.
Thus, knowledge extraction is completed using a combination of available textual data and vector
representations of shape data using Neural Networks.
3.2. Knowledge Graph Construction
By restricting the base node data to product designs, we present a knowledge graph construction method
utilizing three types of nodes: (i) Assemblies, (ii) Subassemblies, and (iii) Parts, in descending order of
hierarchy. The edges or relations between these nodes of the Knowledge Graph are created as follows:
(i) Existing assembly hierarchies obtained from CAD files are used for Assembly-Subassembly-Part
relations, similar to the approach in [54]. This is termed as the Assembly Hierarchy Subgraph, and
it contains 3 types of nodes: part, subassembly, and assembly, which are linked by the relation
type subset_of in hierarchical order.
(ii) The second type of relations between individual parts are created using global similarity of
feature-vectors representing each individual part. The subgraph containing nodes of type part and
similarities between these parts given by the (weighted) relation similar_to is termed the Part
Similarity Subgraph. These global similarity computations are performed using the respective
vector embeddings of the part nodes.
Considering the intended application of this method for large design databases, there are an enormous
number of part-to-part similarity relations that could potentially be created, with a wide range of edge
weights which will affect the goal of efficient search within the KG. Previous work on complex networks
has shown that making networks sparse by removing insignificant relations has the potential for huge
improvements in speed for graph clustering or community detection, while maintaining the original nature
of the graph [55]. This can be achieved for weighted graphs at the global scale using Edge-Weight
Thresholding; work by Yan et al [62] demonstrates the robust nature of network structure when weight
thresholding is applied. In particular, weight-thresholded graphs with a Scale-Free degree distribution
have been shown to maintain the overall topology of the complete graph. Inspired by similar approaches
applied to brain networks in neuroscience as seen in [58] and [59], the construction method is fine-tuned
by constructing a sparse yet meaningful graph through Global Weight Thresholding as described in the
following paragraph.
Let the Part Similarity Subgraph, 𝐺, be described by its symmetric weighted adjacency matrix 𝑊, where
the elements 𝑊𝑎𝑏 denote the weight of the edge connecting nodes a and b. Given a limiting value "α", the
' '
weight thresholded graph 𝐺 then has an adjacency matrix 𝑊 , such that all weights below this value are
'
discarded i.e., if 𝑊𝑎𝑏 < α then 𝑊'𝑎𝑏 = 0. This threshold α is chosen such that the resulting graph 𝐺 is a
scale-free network with a heavy-tailed degree distribution. A scale-free network is one where the fraction
−γ
of nodes P(x) with degree x is close to a power law, i.e., 𝑃(𝑥)∝𝑥 . This phenomenon indicates that there
is a small but significant proportion of nodes with a much higher degree than most other nodes. This
process of simplifying the graph ensures a balance between representativeness and compactness at the
construction level, which speeds up search and detection of structure within the Knowledge Graph.
 Figure 1: Construction Process & Schema of the Design Knowledge Graph

Nodes and relations can be extended where data is available, for instance, adding Finite Element Analysis
(FEA) or Kinematic Simulation data to create localized subgraphs to potentially predict the behavior of
new data. This method provides a starting framework for such additions; the initial relationships are
created considering the KG’s objective of enabling optimal search and inference using a hybrid bottom-up
construction method. Graph data is represented by the Labeled Property Graph model, which allows both
nodes and edges of the graph to be described by key-value pairs which can contain further metadata. This
is advantageous in the case of a design graph, enabling a flexible schema representation for CAD parts
and assemblies, which have different relation types and sets of properties. Fig.1 summarizes the process
of constructing the Product Design Knowledge Graph.
4. Product Design Knowledge Graph: Implementation
4.1. Source Data & Annotation
The KG construction is implemented with data from the FabWave repository described by Bharadwaj et
al [37]. Created as a cyber-infrastructure to automate collection of CAD and related manufacturing data
from academic and open sources, it consists of over 3000 assemblies and over 110,000 parts. This data is
divided the following subsets:
(i) FabWave Categorized: These 4000+ 3D parts were obtained from student-generated data using
design software add-ins and classified into 45 categories, based on both form and process [65].
(ii) FW10C: A 10-class subset from the above dataset used for training classification algorithms.
(iii) FabWave Open: Built from openly available sources of 3D data such as GrabCAD and Autodesk
Fusion Gallery. Consists of 3000+ assemblies, processed into their individual parts (numbering
over 110,000) in multiple data formats (STEP, STL, F3D etc.) along with component tree
structure. Available basic metadata about the parts are recorded during file processing and from
their original sources. [14]
The assemblies in the FabWave dataset were processed into individual parts and organized in the
knowledge graph in accordance with the hierarchical structure of CAD assembly files. This dataset
contains basic assembly categories and size meta-data; however, the individual parts lack detailed
technical labels. The size of the dataset and variety of parts in the FabWave Open dataset means that a
single label classification is a complex exercise and does not necessarily add value to the data. To account
for this, a multi-description scheme was devised, including feature labels, free text descriptions,
process/materials etc. These description categories are covered in Appendix A. With the help of human
annotators with knowledge in design and manufacturing, 3118 of these parts were provided with
descriptions based on this schema, which was combined with shape data to build the knowledge graph.
Annotations are represented as node properties in the KG.
4.2. Knowledge Extraction
As described in section 3.1, global part representations were generated using neural network methods.
Here, we leverage our previous work on a 12-image Multi-View Convolutional Neural Network approach,
specifically modified for CAD data by the addition of dimension measurements for training [28].
Originally used for global classification of 3D shapes, this architecture uses 12 images from multiple
views as input to ResNet CNNs terminating in 4096-dimension vectors, which are then aggregated and
down sampled to a 10-class classifier along with a 1024-dimension vector. The concept of Transfer
Learning enables a fast-paced training process for this neural network, by extending previously trained
models on the ImageNet image classification database to 3D Shape classification and global feature
extraction. Based on the pre-trained 10-label neural network classifier, transferrable features/embeddings
are generated for other all other un-trained 3D models, as demonstrated by Angrish et al. Each of these
embeddings is used as a representative of the corresponding part for calculating global similarity
measures.
4.3. Knowledge Graph Construction
4.3.1. Assembly Hierarchy Subgraph
Our method translates the representation of Assemblies in tree-structure (commonly seen in standard
assembly designs) to a graph. Applying this to the FabWave Open dataset, the Assembly Hierarchy
Subgraph was generated. This subgraph has nodes of type assembly, subassembly and part, in order of
descending hierarchy. The edge type subset_of is used to denote subassembly-assembly and the
part-subassembly hierarchies, thus forming a directed graph.
4.3.2. Part Similarity Subgraph
Further extending the knowledge graph with CNN-based part similarities, the subset of 110,000+
individual part files is considered for similarity computation using their respective vector embeddings.
Each of these parts is represented by a part node and the relations between the nodes are given by an
undirected edge of type similar_to, with the edge weights demoting the cosine similarity (Eqn. 1) between
the Part Embeddings.
 𝑛
∑ 𝐴𝑖 𝐵𝑖
𝐴∙𝐵
𝑐𝑜𝑠 θ = ||𝐴||||𝐵||
= 𝑖=1
(1)
𝑛 𝑛
2 2
∑ 𝐴𝑖 ∑ 𝐵𝑖
𝑖=1 𝑖=1

This measure ranges from 0 (highly dissimilar) to 1 (highly similar) in the positive space. For this dataset,
there exists 12 trillion possible similarity-based connection amongst the parts. Preliminary analysis of
these cosine similarities (Fig. 2(a)) shows that a large majority of these connections denote a weak
similarity between the parts. Constructing the KG using all these relations will result in a graph with a
very high search complexity. However, closer inspection of the frequency distribution of edge weights of
the complete graph in Fig. 2(a) reveals that there are a relatively small but significant number of these
potential connections that denote a high level of part similarity. This phenomenon is likely to appear in
databases with a small but significant fraction of highly similar 3D shapes, as is common in mechanical
design. The interpretation of a “high” similarity level varies between different datasets; this is addressed
by the threshold parameter "α" described in section 3.2. A much lower threshold will result in far more
edges and a highly complex graph, which can provide better search results (at a far higher cost) than a
sparser graph. The threshold parameter is empirically obtained to ensure the scale-free or heavy-tailed
nature of the network, which will ideally ensure a representative yet sparse graph.
Fig. 2(b) shows the observed degree distribution of the thresholded Part Similarity Graph at α = 0.85.
Based on an analysis of this graph’s empirical degree data, a Power-law distribution (γ = 1.808), with an
exponential cutoff is found to be a better fit than the exponential distribution with a log-likelihood ratio R
= 18.1848. This satisfies the minimum condition to exhibit the heavy-tailed nature of the degree
distribution [66], ensuring a compact Scale-Free graph, maintaining the inherent structure of a complete
graph. The hybrid construction method described ultimately results in 27 million relations.

Figure 2: Graph Structure (a) For the Complete Graph: Histogram of Edge Weights: Only a small no. of
edges are relevant; (b) After Weight Thresholding [0.85, 1]: Log-log Fit of the Dist. of Node Degree
shows Exp.-Power-Law Fit

5. Structuring the Knowledge Graph to drive Design Recommendations

5.1. Community Detection in the KG
The most popular approaches to construct KG-based recommendation systems apply manual
segmentation to the KG using classifying schemes, which can limit the results due to the large variety of
possible labels. Recent KG frameworks apply unsupervised methods to group sub-domains in a larger
topic to reduce time to generate and use the knowledge graphs, while still providing accurate
recommendations. Some examples of this approach are seen in the use of local graph structure and
hierarchy-aware systems for predictions [67], [68] and has been applied to fields such as the organization
of scientific publications to aid researchers [69], [70]. In the context of optimal product design search,
community detection within the KG was implemented with the following aims:
(i) Finding Part nodes that are very similar to each other in an unsupervised manner. This similarity
detection reduces the search space.
(ii) Using the internal structure of these communities to simplify searching the KG.
(iii) Provide a framework for probabilistic label recommendations for un-labeled data, using node
similarity within the communities. (Applicable to both Inductive reasoning for new incoming nodes,
as well as Transductive reasoning for existing nodes)
For the thresholded Part Similarity Subgraph of the KG, the underlying structure of the graph is
discovered by finding closely connected communities with highly correlated sets of relations of their
respective nodes. For our application, community detection was applied to a single node type, and can
thus be considered analogous to clustering methods used in data mining [71]. However, these methods
can also be implemented considering multiple relation types if required, which ensures the ability of these
techniques to be extended to more complex data sources. Two important assumptions were made in order
to detect clusters/communities in the Part Similarity Subgraph: (i) Communities are
disjoint/non-overlapping, and (ii) The network does not vary over time.
Extensive reviews of the best community detection/clustering algorithms for graphs are given in [72] and
[73], along with an in-depth performance comparison on benchmarks. Based on these reviews, we
implement Louvain, Leiden, Label Propagation, Walktrap, and Infomap algorithms for Graph Clustering
on the Part Similarity Subgraph of the KG. Importantly, these algorithms do not require prior information
such as the number or size of communities in the network. This makes them well-suited for unsupervised
detection of communities, with the only hyper-parameter being the strength of relations in the graph. This
is an intuitive parameter to control and can be evaluated by the effectiveness of the graph partitions. The
complexity of each of these algorithms are briefly summarized in Table 1, for a graph with 𝑛 nodes, 𝑒
edges and average degree 𝑘. Performance of these algorithms are evaluated using intrinsic measures
described in section 6.
Table 1: Community Detection Algorithms: Complexity

Louvain [74] Has a complexity of 𝑂(𝑒).

Leiden [75] Demonstrated to be faster and to find higher quality clusters than Louvain.
𝑂(𝑛 log 𝑙𝑜𝑔 (𝑘) ) complexity. [76]

Label Propagation [77] Depends on community structure alone; 𝑂(𝑒) complexity.

Walktrap [78] 2
Has 𝑂(𝑛 𝑙𝑜𝑔⁡(𝑛)) complexity for sparse networks.

Infomap [79] Simulates the flow of information in weighted networks. 𝑂(𝑒) complexity.
5.2. Part Retrieval from the KG and Node Centrality
Intelligent retrieval methods using knowledge graphs have been well-studied in multiple domains. For
Knowledge Graph constructed with vector embeddings, we chose an approach that leverages the detected
communities for subsequent greedy nearest-neighbor search. Assuming that communities have been
detected effectively, a single descriptor node for each cluster can massively reduce the complexity of
nearest neighbor search, while maintaining accuracy of search results. We implement this using node
centrality measures within communities. While communities themselves are evaluated using extrinsic
measures, the combined community-retrieval method is evaluated manually using retrieval measures
described in Section 6.
Node-based centrality measures for networks such as Degree, Closeness, Eigenvector, Katz centrality and
PageRank scores have had success for social network analysis and web search results as seen in [80] and
[81]. In particular, the PageRank algorithm [82] has been extensively used in weighted graphs to rank
webpages by order of the most effective at transmitting information to its neighbors. For this design-based
KG, the priority for selecting central nodes is its effectiveness in representing the nodes within its own
community. Using just the network structure to find nodes that have high control over the flow of
information, the PageRank technique was adopted to find central nodes of the communities in the Part
Similarity Subgraph, which were designated as Level-1 Centers. These L-1 center nodes have the highest
centrality scores within each community and are subsequently used to simplify retrieval of data within
KG.
Information retrieval from large databases is a well-studied problem, with the use of Approximate Nearest
Neighbor (k-ANN) Search suitable for high-dimension metric spaces. Well studied approaches to solving
this problem include Space Partitioning/Tree methods (e.g. KD-Trees). Another popular approach to
large-scale ANN tasks is the Proximity-Graph method which generates graphs based on vector similarity
metrics, such as the Navigable Small-World Graphs (NSW) method [83]. By maintaining a proximity
graph with both long-range links as well as short links to closer neighbors, NSW graph methods perform
greedy search with poly-logarithmic complexity by starting at low degree nodes and progressing to higher
degrees. Hierarchical NSW (HNSW) [84] improves this technique to logarithmic search complexity,
separating links by magnitude into multiple layers and performing greedy search on this multi-layer
index. In the case of a similarity-based graph, the lowest layers contain relations with high similarity
weights, with decreasing relation weights in higher levels. Benchmark studies on ANN search
demonstrate that graph-based algorithms such as HNSW outperform Tree and Hashing-based algorithms
at high recall values, while maintaining a relatively compact index size even for large datasets [85].
Leveraging these advantages of the algorithm and the community structure of the KG as described in
section 5.1, the part retrieval technique was implemented as follows:
(i) The HNSW algorithm was used to index the previously defined L-1 Centers/cluster
representatives for Approximate Nearest Neighbor search. (L-1 Centers number 25,706 of the
110,770 total Part nodes, and this reduces the time required to create the index.)
(ii) Vector-embedding were generated for the CAD model of interest to the user, and subsequently
Top-k search was performed on the KG L-1 Centers using the HNSW index.
(iii) This process was then followed by an exhaustive Nearest Neighbors search within the
Communities that the top-k L-1 Centers belong to, thus providing the best matches within the top
candidate communities. The process is visualized in Fig. 3.
The effectiveness of the hybrid ANN and exhaustive search was compared to Exact Nearest Neighbor
Search on all 110,770 Part nodes using the popular KD-Tree method.

Figure 3: Construction of Optimal Retrieval Framework for Part Nodes

6. Evaluating Community Detection and Retrieval

6.1. Community Detection
The quality of Community Detection and Clustering can be evaluated using External or Intrinsic
measures. External measures include Adjusted Rand Index and Normalized Mutual Information (NMI).
However, external measures require knowledge of ground truth communities, which makes them
unsuitable for evaluating our approach which is exploratory in nature. Two intrinsic measures are used in
place of these external metrics: Modularity Density and Silhouette Score, which is another popular
method used to evaluate clustering effectiveness. Modularity Density is a graph specific method which
uses node degree to evaluate graph partitions and is well correlated with the ground truth-based NMI
metric. Both these scores rely on simultaneously finding the level of cohesion within clusters, and the
level of separation between clusters. They are described as follows.
For a community C in a graph partition S, Modularity Density (D) is given by:

1 𝑖𝑛𝑡 𝑜𝑢𝑡
𝐷= 𝑛𝑐
∑ (2λ𝑘𝑖 − 2(1 − λ)𝑘𝑖 ) (2)
𝑖∈𝐶
 𝑖𝑛𝑡
where i is a node within cluster C, 𝑛𝑐 is the number of nodes in cluster C, 𝑘𝑖 is the degree of node i
𝑜𝑢𝑡
within cluster C, and 𝑘𝑖 is the degree of node i external to cluster C. This measure is then averaged over
all clusters to find the score for the entire graph. λ is a tuning parameter; 0.5 is the standard value used.
Silhouette Score for a graph with n nodes is given by:

1 𝑏(𝑖)−𝑎(𝑖)
𝑆= 𝑛
∑ 𝑚𝑎𝑥⁡(𝑎(𝑖), 𝑏(𝑖))
𝑆∈[− 1, 1] (3)
𝑖∈𝑛

Here, for a node i belonging to cluster C, 𝑎(𝑖) is its mean intra-cluster distance (weighted). 𝑏(𝑖) is the
mean (external) weighted distance from i to all nodes in the cluster X, which is the nearest cluster to the
center of C (X ≠ C). The best possible Silhouette Score for a clustering is 1.

Figure 4: Performance of Community Detection Algorithms for Graphs based on (a) Modularity Density
(b) Silhouette Score

Implementing the community discovery algorithms on a range of thresholds from α = 0.85 to α = 0.9,
Label Propagation and Walktrap algorithms at a lower weight threshold level of α = 0.9 shows the best
performance in terms of Modularity Density and Silhouette Score respectively as seen in Fig. 4. The
communities generated by the Label Propagation algorithm were chosen as the best representation of
structure at the lowest level of granularity among parts, since Modularity Density has the highest
correlation to the extrinsic NMI metric for graphs [86]. This algorithm produces 25,706 communities. On
closer examination, 18,055 of these communities contain only a single part. The largest cluster contains
4295 parts; this confirms our previous intuition regarding a significant number of repetitive part shapes
within the database. In summation, through community detection, 7,651 highly similar groups are
identified, incorporating 92,715 out of the 110,770 parts.
Selecting central nodes for each of these communities using the PageRank algorithm gives 25,706 nodes,
one for each at the Level-1 communities. This was considered the first level of the hierarchy in the
knowledge graph or L-1 Centers. This reduces the detection space significantly, allowing for a much
simpler graph representation. The L-1 Centers embeddings are indexed for search using Hierarchical
Navigable Small-World Graphs (HNSW) method for Approximate Nearest Neighbor search.
6.2. Top-k Retrieval Evaluation
Retrieval is evaluated for 29 standard part categories from the FabWave Categorized dataset described in
Section 3.1. Following the method for retrieval described in Section 5.2, an Approximate Nearest
Neighbor (ANN) search on the 25,706 L-1 Center node vectors is used to find top-k clusters, followed by
exhaustive k-Nearest Neighbor (NN) search on the top clusters. This hybrid method is compared with
exhaustive k-NN search on all 110,770 nodes using the popular KD-Tree method. Retrieval is evaluated
using Mean Average Precision (mAP) metric, defined as follows:

( )
𝑘 𝑇𝑃𝑠𝑒𝑒𝑛
1 1
𝑚𝐴𝑃@𝑘 = 𝑛𝑢𝑚_𝑞𝑢𝑒𝑟𝑦 𝐺𝑇𝑃
∑ 𝑖
(4)
𝑖=1

Here, num_query refers to the number of queries made for each class, GTP refers to the total number of
ground truth positives for the query, and TPseen is the number of true positives seen till k. 20 query parts
are evaluated for each class, with each search returning the top-5 neighbors of the queried part. Table 2
shows the performance comparison of the HNSW Index-based Approximate Nearest Neighbor (ANN)
search for knowledge graph part nodes, compared to the KD-Tree based kNN search. Using the advantage
of L-1 centers of the graph communities, the HNSW ANN search index is constructed in far lesser time in
comparison to the KD-Tree method. Searching within the graph using the hybrid HNSW-kNN method is
also achieved in a fraction of the time taken by KD-Tree. This can be attributed to two factors: the
reduction of number of node vectors to index for search, and the decrease in performance of KD-Tree
with high vector size.
As seen in from the results in Table 2, better mAP can be achieved in much lower time using ANN
search, due to a combination of graph community detection and the advantages of the ANN algorithm.
Class-wise mAP values are given in Appendix B, and Appendix C, which shows examples of retrieval for
certain example classes. An analysis of these results show that the ANN approach performs better for
Bushing, Collets, Gasket, Headless Screws, Hex Headed Screws and Washers; in comparison to this
method, the KD-Tree approach performs better for Brackets and Machine Keys. This is attributed to
errors in the design data used for querying (FabWave Categorized); some incomplete part designs in this
dataset lead to substandard part matches for these two classes in both cases. However, the results of
Hybrid ANN + kNN approach are generally better or comparable to that of the baseline KD-Tree method.
These results represent the combined effect of node clustering in the graph and hybrid ANN search,
demonstrating the overall effectiveness of the KG and ANN-based retrieval workflow. In enterprise
databases, this combined method has the potential to account for the tradeoff between speed and accurate
representation of large amounts of data.
Table 2: Comparing the two retrieval methods

Method Search Index Query Time Retrieval: 29 Std. Categories

(seconds)
No. of Constr. mAP@1 mAP@3 mAP@5
Indexed Time (sec)
Nodes
Mean (SD)

HNSW 25,706 17 sec HNSW KNN 0.6913 0.7242 0.7278

followed by
kNN 0.2362 0.0649
(0.0152) (0.1015)
KD-Tree based 110,770 76.61 sec KD-Tree 0.6439 0.6706 0.6760
kNN
2.7859 (0.9022)

6.3. Implementation Details

The 3D CAD model processing and image capture was implemented using Fusion 360 design software
and Blender 3D graphics tool. The Knowledge Graph and all algorithms were implemented on a Windows
10 machine with Intel Xeon W-2225 processor and 64GB RAM. The KG is implemented on Neo4J,
queried with the Cypher language. Visualizations are provided using Graphlytic and Matplotlib. Part and
Assembly metadata are stored on MongoDB, with file storage on Google Drive. Graph Representation
and Clustering were achieved using NetworkX [87], Neo4J’s Graph Data Science extension, CDLib [88],
Infomap [79], Scikit-learn [89], and Scikit-Network [90] packages on Python. The Powerlaw package
[66] was used for testing truncated power-law nature of the degree distribution. hnswlib package [84] is
used for final retrieval using approximate nearest-neighbor (ANN) search. The Fuzzy-Rule system in
section 7.1 is implemented using the Simpful library [91]. Mlxtend [92] was used to implement the
FP-Growth algorithm for recommendation systems shown in Section 7.3.
7. Recommendation Systems using Product Design KG
Search and classification for Product Databases is an active field of study, with multiple potential
applications such as design reuse [32], manufacturer selection [93], function prediction [36], design
recommendation and improving sustainability in design [94]. The effectiveness of the KG for retrieval of
similar components was demonstrated in section 6, which can be used for design reuse by engineers or
Numerically Controlled (NC) process plan reuse. Motivated by such problems, this section shows the
potential to apply recommendation systems to associated problems for engineers, using the advantage of
richer data associativity within the KG.
Retrieving similarly shaped components is one common application for product data stores. Design
efficiency can also be greatly improved by helping designers obtain ‘associated’ standard or non-standard
components, which are likely to be a part of the same subassembly as similar parts of interest. These parts
can be ordered based on the frequency of occurrence in other similar subassemblies, using the part cluster
IDs. Section 7.1 describes a Fuzzy Rule-Based method to obtain these relevant components from the KG.
Using an approach similar to those used in [95], [96] and [49], section 7.2 expands on the use of the KG
for component-based assembly similarity evaluation using shape similarity. Efficient retrieval of CAD
models in large databases consisting of combinations of hundreds of thousands of components is an
industry problem that has primarily relied on PLM/PDM based textual search. Multi-modal search using
3D data in conjunction with associated textual data has the potential to increase effectiveness of the
required suggestions; this is demonstrated in section 7.3.
7.1. Recommending Co-Occurring 3D Designs
Design engineers spend a significant percentage of their time in locating and sorting information, often
comparable to the time spent on solving actual design problems [97]. Recommendation systems can be
powerful tools to improve designers’ productivity; standardized components are often listed in design
software libraries for use by engineers. In the context of new designs, engineers often spend their time
designing critical components which have multiple engineering constraints in an iterative manner. The
KG based recommendation system can improve designer productivity by providing both standard and
non-standard parts associated with high-priority critical components, with multiple variants to choose
from. Components that occur in the same subassemblies as the closest matches to the query part are found
through a combination of ANN search amongst the parts, followed by graph traversal for the Assembly
hierarchy structure in the KG. The parts associated with these Nearest Neighbors (parts in the same
subassembly or assembly) were then analyzed and provided as recommendations to the designer, ranked
by relevance.
From the existing relations of the graph, there are two relations that can be manually inferred:
co_occurring and highly_correlated. Parts that belong to the same subassembly are defined as
co_occurring. Formally, for two parts A and B and a given subassembly SA, one fact that can be inferred
from the graph is:
(𝐴, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑆𝐴)∧(𝐵, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑆𝐴)⟹(𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵)
Parts that belong to the same cluster are defined as highly_correlated. Thus, for parts A & B in the same
cluster,
(𝐴, 𝑠𝑎𝑚𝑒_𝑐𝑙𝑢𝑠𝑡𝑒𝑟, 𝐵)⟹(𝐴, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝐵)
We define the following rule for obtaining potentially relevant co-occurring parts R, for a given query part
Q:
* *
(𝑄, 𝑛𝑒𝑎𝑟𝑒𝑠𝑡_𝑛𝑒𝑖𝑔ℎ𝑏𝑜𝑟, 𝐴) ∧ (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅) ⟹(𝑄, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅)

(𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅)* is inferred through the rule:

* *
(𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵)∧(𝐵, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝑅) ⟹(𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅)
Parts A and B are “transitory” parts obtained from the Knowledge Graph, in the process of finding
relevant parts R. Fig. 5 illustrates a method to use the KG for this application, considering an Unthreaded
Flange as the query part.
 Figure 5: Contextual Design Recommendation Process

The consequence of the above rules is true only to a certain degree; thus, it is denoted by a *. This is due
to the potential inaccuracies induced by the clustering method due to the use of weighted similarity
relations. We approach this rule-based inference problem using a Generalized Fuzzy Rule method, which
has seen success in associated fields, most recently in manufacturing performance measurement [98].
Using these rules, we obtain parts that can potentially be used in conjunction with a given 3D part of
interest. The final measure of truth of the chained rules is calculated using the well-studied Fuzzy
Inference System described by Mamdani and Assilian [99]. This is a linguistic logic modeling technique
where rules are stated using “IF-THEN’ rules, with both the prior and consequent of the rule represented
as fuzzy sets. The degree of membership of each element in the fuzzy set is determined by Membership
Functions (range of 0 to 1) heuristically obtained from source data.
 Figure 6: (a) Distribution of Silhouette Score in the KG (b) Membership Func. for “highly_correlated”
(Low, Medium and High)

The uncertainty is introduced in the rule-based system through the component (𝐵, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝑅
). “highly_correlated” is based on parts belonging to the same cluster; based on the results in section 6.2,
the effectiveness of this property can be measured using the Silhouette Score of the cluster containing B
and R. Since a lower silhouette score in the range [-1, 1] indicates a low or negative correlation between
the parts, 3 different fuzzy sets are utilized to describe the property “highly_correlated” denoted as Low,
Medium or High. Based on prior knowledge of the silhouette score as seen in Fig. 6, the sets are defined
with the following membership functions: Sigmoid (Low) and Trapezoidal (for Medium and High). The
property “co_occurring” is represented by High or Low; since it is a fact derived from the graph in
(𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵), this would correspond to a True/False value. Based on these definitions, the
linguistic rules for the fuzzy inference system are thus re-defined as follows:
𝐼𝐹 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵) 𝐼𝑆 𝐻𝑖𝑔ℎ 𝐴𝑁𝐷 (𝐵, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝑅) 𝐼𝑆 𝐻𝑖𝑔ℎ 𝑇𝐻𝐸𝑁 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅) 𝐼𝑆 𝐻𝑖𝑔ℎ
𝐼𝐹 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵) 𝐼𝑆 𝐻𝑖𝑔ℎ 𝐴𝑁𝐷 (𝐵, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝑅) 𝐼𝑆 𝑀𝑒𝑑𝑖𝑢𝑚 𝑇𝐻𝐸𝑁 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅) 𝐼𝑆 𝐻𝑖𝑔ℎ

𝐼𝐹 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵) 𝐼𝑆 𝐻𝑖𝑔ℎ 𝐴𝑁𝐷 (𝐵, ℎ𝑖𝑔ℎ𝑙𝑦_𝑐𝑜𝑟𝑟𝑒𝑙𝑎𝑡𝑒𝑑, 𝑅) 𝐼𝑆 𝐿𝑜𝑤 𝑇𝐻𝐸𝑁 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅) 𝐼𝑆 𝐿𝑜𝑤

𝐼𝐹 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝐵) 𝐼𝑆 𝐿𝑜𝑤 𝑇𝐻𝐸𝑁 (𝐴, 𝑐𝑜_𝑜𝑐𝑐𝑢𝑟𝑟𝑖𝑛𝑔, 𝑅) 𝐼𝑆 𝐿𝑜𝑤

 Figure 7: Recommended Co-occurring Parts with their corresponding Truth Values

The final truth value is obtained from the conjunction of rules as shown previously in this section, and
denotes its actual relevance based on the structure of the knowledge graph. Fig. 7 shows results of the
fuzzy inference system with relevance scores. A manual check of the results shows a few relevant
co-occurring parts for 3 example queries, which are circled in blue. Since the truth values of the
recommendations are based on the graph structure (which is not perfect), the ordering and accuracy of the
results is not ideal; further investigation can reveal more complex and accurate rule-based systems for part
recommendation from the knowledge graph.
7.2. Detecting Similarity of Assemblies
To compute similarity of assemblies, the clusters of their constituent parts in the KG can be used as a
factor. Every assembly was represented by a multiset containing each of its constituent parts’ community
IDs at Level-1. The rule for similarity between two assemblies X & Y can thus be defined as follows:
(𝑝, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑋) ⟹∃𝐴(𝐴, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑋)

(𝑞, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑌) ⟹∃𝐵(𝐵, 𝑠𝑢𝑏𝑠𝑒𝑡_𝑜𝑓, 𝑌)

∃𝐴∃𝐵 (𝐴, 𝑠𝑎𝑚𝑒_𝑐𝑙𝑢𝑠𝑡𝑒𝑟, 𝐵)⟹(𝑋, 𝑠𝑖𝑚𝑖𝑙𝑎𝑟_𝑡𝑜, 𝑌)*

The similar_to relation in the above rule between these assemblies X & Y was quantified using the Jaccard
index for the similarity of sets. A modified version of the Jaccard Index (Eqn. 4) which rewards repeated
values in a multiset was used:
|𝑋𝑋∩𝑌𝑌|
𝐽(𝑋, 𝑌) = |𝑋𝑋|+|𝑌𝑌|
𝐽∈[0, 0. 5] (5)

where XX and YY are the multiset collection of part L-1 cluster IDs for two assemblies X & Y
respectively.
 Figure 8: Clusters of Assemblies based on individual part similarity

This similarity measure can be used to compute clusters among assemblies, based on the similarities of
their individual parts. Fig. 8 shows some examples of community structure for assemblies. This can easily
be extended to sub-assembly similarity search by the same method. The practical implication of building
this data in the KG is future improvement in modular product design, enabling flexible production and
low volume customization. Using these advantages, the best practices previously followed for
design-for-assembly and design-for-manufacture can be adopted by designers and further refined to
ultimately create better products. The flexibility of the KG approach can be used to include data
pertaining to electromechanical or hydraulic systems, which can provide valuable information to
designers and simplify communication across domains, with types of node clusters detected based on
multiple relation types.
7.3. Design and Manufacturing Parameter Recommendations
As described in Section 3.1, the Knowledge Graph contains annotations from users on various design and
manufacturing parameters for parts in the database. Although these annotations cover only a fraction of
the KG nodes, the concept of collaborative filtering can be applied based on the clusters of parts detected
within the KG. Collaborative Filtering (CF) is a popular method used in web-based recommender systems
to provide fast and accurate recommendations to users, based on the preferences of similar users who
have previously interacted with the system. In the context of this KG, the recommendations provided
depend on the subset of previously annotated nodes which are the most similar to a search node. Using a
combination of the clusters in the graph and pattern mining, a method that can be described as
Clustering-Model Based Collaborative Filtering technique [100] was used to provide recommendations.
 Figure 9: Recommendations for 6 parts: Recommendations for Spring and Triangular component are
provided using Multimodal search using 3D Data + Text

Based on the annotations for the parts given by users described in Appendix A, design and manufacturing
parameter recommendations are generated using the Frequent-Pattern Tree Growth (FP-Tree) approach
[101]. This approach decomposes pattern mining into smaller sub-tasks for more efficient search and is
faster than the well-known Apriori algorithm for the same task. For each part node that has been
annotated, the set of individual descriptions are considered for parameter search. This algorithm evaluates
the relevance of parameter recommendations using the probability of co-occurring parameters, given
primarily by two parameters: Support and Lift. Support is the measure of how often a parameter A occurs
in the relevant nodes of the KG i.e, its frequency of occurrence in the set of node parameters. Lift is an
extension of this parameter, which describes whether two parameter sets X and Y are independent. It is
given by Eqn. 5:
𝑆𝑢𝑝𝑝𝑜𝑟𝑡(𝑋∪𝑌)
𝐿𝑖𝑓𝑡(𝑋⟹𝑌) = 𝑆𝑢𝑝𝑝𝑜𝑟𝑡(𝑋)×𝑆𝑢𝑝𝑝𝑜𝑟𝑡(𝑌)
(6)

A Lift value of 1 suggests that the parameter sets X and Y are independent of each other. Higher the lift
value, the stronger the degree of dependence of the parameter sets X & Y.
The recommended parameters for the Unthreaded Flange include the feature commonly seen in associated
parts in the database (‘Holes’), along with its material and process suggestions (‘Metal Alloys’,
‘Casting’). Fig. 9 shows the recommendations for multiple parts. Parameters for the Spring component are
queried with a part category (‘Automobile’) along with shape and dimensions, providing parameter
recommendations relevant to automotive spring components. Similarly, the triangular component is
queried with a part description denoting ‘Industrial Component’. These results are dependent on the
amount of data available; currently, the annotations are sparse relative to the size of data. Further
investigation on graph-recommendation systems can yield superior recommendations.
8. Discussion and Future Work
Structuring product data in clusters reveals the nature of this design data store created from open sources:
a mix of unique designs and a significant fraction of repeated non-standard parts, along with standardized
parts. Similar data in enterprise siloes can serve an important purpose for future designs: either through
design reuse or by enabling data-driven parameter suggestions, in contrast to manual retrieval by
text-based search or advanced image only visual search. With minimal human intervention, existing
patterns detected within the corpus of designs provide a rich source of engineering decisions. The ability
of an automatically structured KG for efficiently implementing recommendations is demonstrated, despite
minimal availability of annotated data.
We note a few potential weaknesses to the KG construction approach and the data used:
(i) Local relation density differences in large KGs are not considered for structure detection and may
lead to issues with accuracy and scalability.
(ii) The assumption of disjoint communities, and the static nature of these communities will not hold
with extensive additions to the database over time.
(iii) The current approach uses shape similarity alone to both prune relations and detect local structure
in the graph. This method does not currently consider other relation types that can potentially be
created, particularly with the significant amount of associated data with product models available
within enterprise stores.
(iv) The relative lack of rich annotations or textual descriptions in our dataset limits the types of
relations and recommendations that can be provided to designers.
Future improvements to the data must focus on improved representations of parts through a combination
of automated feature/process recognition, and connecting existing data to design documents,
manufacturing parameters, simulations and other geometric and textual part data. With improved data
connectivity, the KG can be used for more complex KG-based recommendation applications such as
mining design rule updates from downstream manufacturing and quality inspection data, similar to the
work by Ko et al [102].
The construction of the KG has a large bearing on the quality of recommendations that can be provided
using this technique. For the cosine similarity relations created using a bottom-up method, the
thresholding parameter described in section 3.2 and implemented in 4.3.2 affects the quality of edges in
the Part Similarity Subgraph. A lower threshold could provide better recommendations at the cost of
greater complexity. However, if unweighted methods are used for graph traversal and recommendation
systems, it is possible that low-quality recommendations could be obtained. In this semi-supervised
framework, it is thus critical to choose a threshold that is suitable for downstream tasks. An incremental
edge addition method that optimizes search results is an area for future exploration.
The assumption of disjoint and static communities relies on the large size of the database relative to new
additions. Future additions to the knowledge graph require these issues to be addressed to ensure scalable
solutions, with dynamically responsive KG structuring methods. The problem of localized graph structure
variation is visualized in Fig. 10. Basic inspection reveals that while there are very dense clusters in the
graph with high internal part similarity, there are also multiple clusters that can be merged. There are also
some relatively large clusters that can be divided into distinct subsets. Fine-tuning these clusters can be
achieved by Incremental Cluster Repairing of existing clustered similarity graphs, using newer methods
such as n-depth re-clustering as demonstrated by Saeedi et al [103]. However, this is not an issue with
effectiveness of clustering alone; the exclusive use of similarity relations for clustering is another
contributor to variability. Using multiple relations and node features to find structure in the KG is the next
area of focus to improve the KG. The interrelated aspects of clustering efficiency and encoding multiple
relation types can be addressed by graph construction techniques that consider the dynamically varying
nature of different neighborhoods and multiple-relation types of the graph, as seen in the work by
Fakhraei et al [104]. Alternately, the graph structure can be encoded in rule-based learning or graph
embedding methods to combine structuring and reasoning.

Figure 10: Visual analysis of some clusters

The size of enterprise data is often several orders of magnitude larger than the data corpus used in this
paper; scalable approaches for multi-modal search are an important priority. Future research directions in
intelligent product design using KGs necessitate larger and more complex graphs. The Approximate
Nearest Neighbor and Collaborative Filtering methods used for reasoning over the KG have limitations:
the ANN method currently indexes only a single relation type, and collaborative filtering suffers from
cold-start issues with prediction due to data sparsity. The techniques must eventually be replaced by
multi-hop inference using latest knowledge reasoning methods, most popular of them being Graph
Embeddings. While embeddings work well on large-scale graphs, they tend to lack interpretability. To
create recommendation systems in the design domain, Inductive Subgraph Reasoning methods show more
promise, since they learn the underlying relational semantics of the graph. Work by Teru et al [105]
demonstrates the strong performance of this technique in comparison to embedding based methods on
standard benchmarks. In the future, we plan to investigate similar methods for more complex
decision-making problems, focused on design parameter recommendation, manufacturer search and data
driven automation.
9. Conclusions
In this paper, we present a novel method of constructing Product Design Knowledge Graphs (KG) with
3D product designs at the core, to enable intelligent design recommendations. Our contributions are
summed up as follows:
(i) We demonstrate a Hybrid Knowledge Graph construction methodology where Top-Down
relations are constructed using hierarchical assembly-part relations obtained from technical
descriptions of designs. Subsequently, an efficient technique for Bottom-Up construction of Part
Similarity relations is demonstrated, by applying Neural-Network generated Shape Vectors and
Global Weight Thresholding to create a sparse yet representative graph.
(ii) Using our dataset collected from openly available sources, we demonstrate the effectiveness of a
KG constructed by this method for recommendation of similar shapes. This is implemented using
a combination of Unsupervised Graph Community Detection and Approximate Nearest Neighbor
search on the knowledge graph.
(iii) We illustrate three potential applications of this method: (a) Rule-Based retrieval of potentially
associated components, (b) Similarity Detection for Assemblies, and (c) Contextually relevant
design parameter recommendation using Collaborative Filtering.
Similar approaches can be implemented for multi-modal search in enterprise databases, combining both
3D components and textual data obtained from prior design-associated tasks. The addition of further
product lifecycle data obtained from diverse enterprise data sources have the potential to enable
multi-domain decision making. Future work aims to improve the 3D-driven KG recommendation systems
in design and manufacturing using link prediction and inductive reasoning methods.
Acknowledgement
This research was supported in part by NSF Grant#1937043. The authors do not have any competing
interests with respect to the work described in this paper.
References
[1] T. Vernica, A. Hanke, and W. Bernstein, “Leveraging standard geospatial representations for
industrial augmented reality,” National Institute of Standards and Technology, Gaithersburg, MD,
NIST AMS 100-29, Apr. 2020. doi: 10.6028/NIST.AMS.100-29.
[2] J. Szarazi and C. Bock, “Machine-Readable Physics to Improve Collaboration and Process
Management for Design Simulation,” National Institute of Standards and Technology, Gaithersburg,
MD, NIST AMS 100-29, Apr. 2020. doi: 10.6028/NIST.AMS.100-29.
[3] R. Geissbauer, J. Wunderlin, S. Schrauf, J. H. Krause, J.-T. Morr, and A. Odenkirchen, “Digital
Product Development 2025.” PricewaterhouseCoopers GmbH, Mar. 2019. [Online]. Available:
https://www.pwc.de/en/digitale-transformation/digital-product-development-2025.html
[4] S. K. Sim and A. H. Duffy, “A foundation for machine learning in design,” AI EDAM, vol. 12, no. 2,
pp. 193–209, 1998.
[5] S. K. Chandrasegaran et al., “The evolution, challenges, and future of knowledge representation in
product design systems,” Comput.-Aided Des., vol. 45, no. 2, pp. 204–228, 2013.
[6] A. K. Goel, S. Vattam, B. Wiltgen, and M. Helms, “Cognitive, collaborative, conceptual and
creative—four characteristics of the next generation of knowledge-based CAD systems: a study in
biologically inspired design,” Comput.-Aided Des., vol. 44, no. 10, pp. 879–900, 2012.
[7] I. Gibson, D. Rosen, B. Stucker, and M. Khorasani, “Design for Additive Manufacturing,” in
Additive Manufacturing Technologies, Cham: Springer International Publishing, 2021, pp. 555–607.
doi: 10.1007/978-3-030-56127-7_19.
[8] X. L. Dong et al., “AutoKnow: Self-Driving Knowledge Collection for Products of Thousands of
Types,” in Proceedings of the 26th ACM SIGKDD International Conference on Knowledge
Discovery & Data Mining, Virtual Event CA USA, Aug. 2020, pp. 2724–2734. doi:
10.1145/3394486.3403323.
[9] C. Meilicke, M. W. Chekol, D. Ruffinelli, and H. Stuckenschmidt, “Anytime Bottom-Up Rule
Learning for Knowledge Graph Completion.,” in Proceedings of the Twenty-Eighth International
Joint Conference on Artificial Intelligence (IJCAI), 2019, pp. 3137–3143.
[10] A. Bosselut, H. Rashkin, M. Sap, C. Malaviya, A. Celikyilmaz, and Y. Choi, “COMET:
Commonsense Transformers for Automatic Knowledge Graph Construction,” ArXiv190605317 Cs,
Jun. 2019, Accessed: Sep. 25, 2021. [Online]. Available: http://arxiv.org/abs/1906.05317
[11] P. Z. Chądzyński and V. McQueen, “The MBE Vision needs MBD to reach outside its current
MCAD and PMI comfort zone,” Model-Based Enterp. Summit MBE 2020, p. 70, 2020.
[12] X. Li, M. Lyu, Z. Wang, C.-H. Chen, and P. Zheng, “Exploiting knowledge graphs in industrial
products and services: A survey of key aspects, challenges, and future perspectives,” Comput. Ind.,
vol. 129, p. 103449, Aug. 2021, doi: 10.1016/j.compind.2021.103449.
[13] Bharadwaj, Akshay Ganesh and Starly, Binil, “FabWave Product Design Knowledge Graph
(FPD-KG).” Zenodo, Feb. 14, 2022. doi: 10.5281/ZENODO.6083697.
[14] B. Starly, Bharadwaj, Akshay Ganesh, and A. Angrish, “Fabwave-3D Part Repository and Product
Design Knowledge Graph,” Fabwave- 3D Part Repository.
https://www.dimelab.org/knowledge-graphs (accessed May 18, 2022).
[15] T. P. Tanon, G. Weikum, and F. Suchanek, “Yago 4: A reason-able knowledge base,” in The
Semantic Web. ESWC 2020. Lecture Notes in Computer Science, 2020, vol. 12123, pp. 583–596.
doi: https://doi.org/10.1007/978-3-030-49461-2_34.
[16] D. Vrandečić and M. Krötzsch, “Wikidata: a free collaborative knowledgebase,” Commun. ACM,
vol. 57, no. 10, pp. 78–85, 2014.
[17] A. Bordes, N. Usunier, A. Garcia-Duran, J. Weston, and O. Yakhnenko, “Translating embeddings
for modeling multi-relational data,” Adv. Neural Inf. Process. Syst., vol. 26, 2013.
[18] J. Yuan et al., “Constructing biomedical domain-specific knowledge graph with minimum
supervision,” Knowl. Inf. Syst., vol. 62, no. 1, pp. 317–336, Jan. 2020, doi:
10.1007/s10115-019-01351-4.
[19] J. Dörpinghaus, A. Stefan, B. Schultz, and M. Jacobs, “Towards context in large scale biomedical
knowledge graphs,” ArXiv200108392 Cs, Jan. 2020, Accessed: Jun. 22, 2021. [Online]. Available:
http://arxiv.org/abs/2001.08392
[20] D. Mrdjenovich et al., “propnet: A Knowledge Graph for Materials Science,” Matter, vol. 2, no. 2,
pp. 464–480, Feb. 2020, doi: 10.1016/j.matt.2019.11.013.
[21] B. Zhou, J. Bao, Y. Liu, and D. Song, “BA-IKG: BiLSTM Embedded ALBERT for Industrial
Knowledge Graph Generation and Reuse,” in 2020 IEEE 18th International Conference on
Industrial Informatics (INDIN), Warwick, United Kingdom, Jul. 2020, pp. 63–69. doi:
10.1109/INDIN45582.2020.9442198.
[22] E. B. Myklebust, E. Jimenez-Ruiz, J. Chen, R. Wolf, and K. E. Tollefsen, “Knowledge Graph
Embedding for Ecotoxicological Effect Prediction,” ArXiv190701328 Cs, vol. 11779, pp. 490–506,
2019, doi: 10.1007/978-3-030-30796-7_30.
[23] Z. Wu et al., “3d shapenets: A deep representation for volumetric shapes,” in Proceedings of the
IEEE conference on computer vision and pattern recognition, 2015, pp. 1912–1920.
[24] D. Maturana and S. Scherer, “Voxnet: A 3d convolutional neural network for real-time object
recognition,” in 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), 2015, pp. 922–928. doi: 10.1109/IROS.2015.7353481.
[25] C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification
and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern
recognition, 2017, pp. 652–660.
[26] H. Su, S. Maji, E. Kalogerakis, and E. Learned-Miller, “Multi-view convolutional neural networks
for 3d shape recognition,” in Proceedings of the IEEE international conference on computer vision,
2015, pp. 945–953.
[27] A. Kanezaki, Y. Matsushita, and Y. Nishida, “Rotationnet: Joint object categorization and pose
estimation using multiviews from unsupervised viewpoints,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition, 2018, pp. 5010–5019.
[28] A. Angrish, A. Bharadwaj, and B. Starly, “MVCNN++: Computer-Aided Design Model Shape
Classification and Retrieval Using Multi-View Convolutional Neural Networks,” J. Comput. Inf.
Sci. Eng., vol. 21, no. 1, p. 011001, Feb. 2021, doi: 10.1115/1.4047486.
[29] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “Imagenet: A large-scale hierarchical
image database,” in 2009 IEEE conference on computer vision and pattern recognition, 2009, pp.
248–255.
[30] Z. Zhang, P. Jaiswal, and R. Rai, “Featurenet: Machining feature recognition based on 3d
convolution neural network,” Comput.-Aided Des., vol. 101, pp. 12–22, 2018.
[31] Y. Song, F. He, Y. Duan, Y. Liang, and X. Yan, “A Kernel Correlation-Based Approach to
Adaptively Acquire Local Features for Learning 3D Point Clouds,” Comput.-Aided Des., vol. 146,
p. 103196, 2022.
[32] E. Clark, A. Vincent, J. N. Kutz, and S. L. Brunton, “Bracketing brackets with bras and kets,” J.
Manuf. Syst., vol. 58, pp. 384–391, Jan. 2021, doi: 10.1016/j.jmsy.2020.12.018.
[33] S. Jayanti, Y. Kalyanaraman, N. Iyer, and K. Ramani, “Developing an engineering shape benchmark
for CAD models,” Comput.-Aided Des., vol. 38, no. 9, pp. 939–953, Sep. 2006, doi:
10.1016/j.cad.2006.06.007.
[34] S. Koch et al., “Abc: A big cad model dataset for geometric deep learning,” in Proceedings of the
IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 9601–9611.
[35] S. Kim, H. Chi, X. Hu, Q. Huang, and K. Ramani, “A large-scale annotated mechanical components
benchmark for classification and retrieval tasks with deep neural networks,” in Computer
Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings,
Part XVIII 16, 2020, pp. 175–191.
[36] V. Ferrero, K. Hassani, D. Grandi, and B. DuPont, “Classifying Component Function in Product
Assemblies with Graph Neural Networks,” ArXiv210707042 Cs, Jul. 2021, Accessed: Sep. 14,
2021. [Online]. Available: http://arxiv.org/abs/2107.07042
[37] A. Bharadwaj, Y. Xu, A. Angrish, Y. Chen, and B. Starly, “Development of a pilot manufacturing
cyberinfrastructure with an information rich mechanical CAD 3D model repository,” in
International Manufacturing Science and Engineering Conference, 2019, vol. 58745, p.
V001T02A035.
[38] K.-Y. Kim, D. G. Manley, and H. Yang, “Ontology-based assembly design and information sharing
for collaborative product development,” Comput.-Aided Des., vol. 38, no. 12, pp. 1233–1250, Dec.
2006, doi: 10.1016/j.cad.2006.08.004.
[39] L. Qiao, Y. Qie, Z. Zhu, Y. Zhu, U. K. uz Zaman, and N. Anwer, “An ontology-based modelling and
reasoning framework for assembly sequence planning,” Int. J. Adv. Manuf. Technol., vol. 94, no.
9–12, pp. 4187–4197, Feb. 2018, doi: 10.1007/s00170-017-1077-4.
[40] Z. Ming, G. Sharma, J. K. Allen, and F. Mistree, “An Ontology for Representing Knowledge of
Decision Interactions in Decision-Based Design,” Comput. Ind., vol. 114, p. 103145, Jan. 2020, doi:
10.1016/j.compind.2019.103145.
[41] N. Wan, R. Mo, L. Liu, and J. Li, “New methods of creating MBD process model: On the basis of
machining knowledge,” Comput. Ind., vol. 65, no. 4, pp. 537–549, May 2014, doi:
10.1016/j.compind.2013.12.005.
[42] Y. Zhong, C. Jiang, Y. Qin, G. Yang, M. Huang, and X. Luo, “Automatically generating assembly
sequences with an ontology-based approach,” Assem. Autom., vol. 40, no. 2, pp. 319–334, Nov.
2019, doi: 10.1108/AA-12-2018-0271.
[43] R. Wang et al., “Ontology-Based Representation of Meta-Design in Designing Decision
Workflows,” J. Comput. Inf. Sci. Eng., vol. 19, no. 1, p. 011003, Mar. 2019, doi:
10.1115/1.4041474.
[44] F. Boussuge et al., “Capturing simulation intent in an ontology: CAD and CAE integration
application,” J. Eng. Des., vol. 30, no. 10–12, pp. 688–725, Dec. 2019, doi:
10.1080/09544828.2019.1630806.
[45] H. Cheong and A. Butscher, “Physics-based simulation ontology: an ontology to support modelling
and reuse of data for physics-based simulation,” J. Eng. Des., vol. 30, no. 10–12, pp. 655–687, Dec.
2019, doi: 10.1080/09544828.2019.1644301.
[46] Z. Wu et al., “Semantic hyper-graph-based knowledge representation architecture for complex
product development,” Comput. Ind., vol. 100, pp. 43–56, Sep. 2018, doi:
10.1016/j.compind.2018.04.008.
[47] X. Li, S. Zhang, R. Huang, B. Huang, C. Xu, and B. Kuang, “Structured modeling of heterogeneous
CAM model based on process knowledge graph,” Int. J. Adv. Manuf. Technol., vol. 96, no. 9–12, pp.
4173–4193, Jun. 2018, doi: 10.1007/s00170-018-1862-8.
[48] Z. Han, R. Mo, H. Yang, and L. Hao, “CAD assembly model retrieval based on multi-source
semantics information and weighted bipartite graph,” Comput. Ind., vol. 96, pp. 54–65, Apr. 2018,
doi: 10.1016/j.compind.2018.01.003.
[49] K. Lupinetti, F. Giannini, M. Monti, and J.-P. Pernot, “Content-based multi-criteria similarity
assessment of CAD assembly models,” Comput. Ind., vol. 112, p. 103111, Nov. 2019, doi:
10.1016/j.compind.2019.07.001.
[50] C. Zhang, G. Zhou, Q. Lu, and F. Chang, “Graph-based knowledge reuse for supporting
knowledge-driven decision-making in new product development,” Int. J. Prod. Res., vol. 55, no. 23,
pp. 7187–7203, Dec. 2017, doi: 10.1080/00207543.2017.1351643.
[51] T. D. Hedberg Jr, M. Bajaj, and J. A. Camelio, “Using graphs to link data across the product
lifecycle for enabling smart manufacturing digital threads,” J. Comput. Inf. Sci. Eng., vol. 20, no. 1,
p. 011011, 2020.
[52] J. Martinez-Gil et al., “General Model for Tracking Manufacturing Products Using Graph
Databases,” in Data-Driven Process Discovery and Analysis, vol. 379, P. Ceravolo, M. van Keulen,
and M. T. Gómez-López, Eds. Cham: Springer International Publishing, 2020, pp. 86–100. doi:
10.1007/978-3-030-46633-6_5.
[53] W. Nie, Y. Wang, D. Song, and W. Li, “3D Model Retrieval Based on a 3D Shape Knowledge
Graph,” IEEE Access, vol. 8, pp. 142632–142641, 2020, doi: 10.1109/ACCESS.2020.3013595.
[54] B. Zhou, J. Bao, Z. Chen, and Y. Liu, “KGAssembly: Knowledge graph-driven assembly process
generation and evaluation for complex components,” Int. J. Comput. Integr. Manuf., pp. 1–21, Mar.
2021, doi: 10.1080/0951192X.2021.1891572.
[55] G. Buchgeher, D. Gabauer, J. Martinez-Gil, and L. Ehrlinger, “Knowledge Graphs in Manufacturing
and Production: A Systematic Literature Review,” ArXiv201209049 Cs, Dec. 2020, Accessed: Jun.
22, 2021. [Online]. Available: http://arxiv.org/abs/2012.09049
[56] R. Pinquié, P. Véron, F. Segonds, and T. Zynda, “A Property Graph Data Model for a
Context-Aware Design Assistant,” in Product Lifecycle Management in the Digital Twin Era, Cham,
2019, pp. 181–190.
[57] A. J. Donkers, D. Yang, and N. H. Baken, “Linked Data for Smart Homes: Comparing RDF and
Labeled Property Graphs,” in Proceedings of the 8th Linked Data in Architecture and Construction
Workshop, Dublin, Ireland, Jun. 2020, pp. 23–36.
[58] A. Huet, F. Segonds, R. Pinquie, P. Veron, J. Guegan, and A. Mallet, “Context-aware cognitive
design assistant: Implementation and study of design rules recommendations,” Adv. Eng. Inform.,
vol. 50, p. 101419, Oct. 2021, doi: 10.1016/j.aei.2021.101419.
[59] J. Hao, L. Zhao, J. Milisavljevic-Syed, and Z. Ming, “Integrating and navigating engineering design
decision-related knowledge using decision knowledge graph,” Adv. Eng. Inform., vol. 50, p. 101366,
Oct. 2021, doi: 10.1016/j.aei.2021.101366.
[60] Z. Zhao, S.-K. Han, and I.-M. So, “Architecture of knowledge graph construction techniques,” Int.
J. Pure Appl. Math., vol. 118, no. 19, pp. 1869–1883, 2018.
[61] Y. Feng, Z. Zhang, X. Zhao, R. Ji, and Y. Gao, “GVCNN: Group-view convolutional neural
networks for 3D shape recognition,” in Proceedings of the IEEE Conference on Computer Vision
and Pattern Recognition, 2018, pp. 264–272.
[62] X. Yan, L. G. S. Jeub, A. Flammini, F. Radicchi, and S. Fortunato, “Weight Thresholding on
Complex Networks,” ArXiv180607479 Phys., Oct. 2018, doi: 10.1103/PhysRevE.98.042304.
[63] C. R. Buchanan et al., “The effect of network thresholding and weighting on structural brain
networks in the UK Biobank,” NeuroImage, vol. 211, p. 116443, May 2020, doi:
10.1016/j.neuroimage.2019.116443.
[64] O. Civier, R. E. Smith, C.-H. Yeh, A. Connelly, and F. Calamante, “Is removal of weak connections
necessary for graph-theoretical analysis of dense weighted structural connectomes from diffusion
MRI?,” NeuroImage, vol. 194, pp. 68–81, Jul. 2019, doi: 10.1016/j.neuroimage.2019.02.039.
[65] B. Starly, A. Bharadwaj, and A. Angrish, “FabWave CAD Repository Categorized Part Classes -
CAD 1 through 15 Classes (Part 1/3).” Dec. 04, 2019. doi: 10.13140/RG.2.2.31167.87201.
[66] J. Alstott, E. Bullmore, and D. Plenz, “Powerlaw: a Python package for analysis of heavy-tailed
distributions,” PLoS ONE, vol. 9, no. 1, p. e85777, Jan. 2014, doi: 10.1371/journal.pone.0085777.
[67] Z. Yang and S. Dong, “HAGERec: Hierarchical Attention Graph Convolutional Network
Incorporating Knowledge Graph for Explainable Recommendation,” Knowl.-Based Syst., vol. 204,
p. 106194, Sep. 2020, doi: 10.1016/j.knosys.2020.106194.
[68] Z. Zhang, F. Zhuang, M. Qu, F. Lin, and Q. He, “Knowledge graph embedding with hierarchical
relation structure,” in Proceedings of the 2018 Conference on Empirical Methods in Natural
Language Processing, 2018, pp. 3198–3207.
[69] C. Carusi and G. Bianchi, “Scientific community detection via bipartite scholar/journal graph
co-clustering,” J. Informetr., vol. 13, no. 1, pp. 354–386, Feb. 2019, doi: 10.1016/j.joi.2019.01.004.
[70] M. D. L. Tosi and J. C. dos Reis, “SciKGraph: A knowledge graph approach to structure a scientific
field,” J. Informetr., vol. 15, no. 1, p. 101109, Feb. 2021, doi: 10.1016/j.joi.2020.101109.
[71] R. Guidotti and M. Coscia, “On the Equivalence Between Community Discovery and Clustering,”
in Smart Objects and Technologies for Social Good, vol. 233, B. Guidi, L. Ricci, C. Calafate, O.
Gaggi, and J. Marquez-Barja, Eds. Cham: Springer International Publishing, 2018, pp. 342–352.
doi: 10.1007/978-3-319-76111-4_34.
[72] M. A. Javed, M. S. Younis, S. Latif, J. Qadir, and A. Baig, “Community detection in networks: A
multidisciplinary review,” J. Netw. Comput. Appl., vol. 108, pp. 87–111, Apr. 2018, doi:
10.1016/j.jnca.2018.02.011.
[73] Z. Yang, R. Algesheimer, and C. J. Tessone, “A Comparative Analysis of Community Detection
Algorithms on Artificial Networks,” Sci. Rep., vol. 6, no. 1, p. 30750, Aug. 2016, doi:
10.1038/srep30750.
[74] V. D. Blondel, J.-L. Guillaume, R. Lambiotte, and E. Lefebvre, “Fast unfolding of communities in
large networks,” J. Stat. Mech. Theory Exp., vol. 2008, no. 10, p. P10008, 2008.
[75] V. Traag, L. Waltman, and N. J. van Eck, “From Louvain to Leiden: guaranteeing well-connected
communities,” Sci. Rep., vol. 9, no. 1, p. 5233, Dec. 2019, doi: 10.1038/s41598-019-41695-z.
[76] V. A. Traag, “Faster unfolding of communities: Speeding up the Louvain algorithm,” Phys. Rev. E,
vol. 92, no. 3, p. 032801, 2015.
[77] U. N. Raghavan, R. Albert, and S. Kumara, “Near linear time algorithm to detect community
structures in large-scale networks,” Phys. Rev. E, vol. 76, no. 3, p. 036106, 2007.
[78] P. Pons and M. Latapy, “Computing communities in large networks using random walks,” presented
at the J. Graph Algorithms Appl, 2006.
[79] M. Rosvall and C. T. Bergstrom, “Maps of random walks on complex networks reveal community
structure,” Proc. Natl. Acad. Sci., vol. 105, no. 4, pp. 1118–1123, 2008.
[80] A. Landherr, B. Friedl, and J. Heidemann, “A critical review of centrality measures in social
networks,” Bus. Inf. Syst. Eng., vol. 2, no. 6, pp. 371–385, 2010.
[81] F. A. Rodrigues, “Network Centrality: An Introduction,” in A Mathematical Modeling Approach
from Nonlinear Dynamics to Complex Systems, vol. 22, E. E. N. Macau, Ed. Cham: Springer
International Publishing, 2019, pp. 177–196. doi: 10.1007/978-3-319-78512-7_10.
[82] L. Page, S. Brin, R. Motwani, and T. Winograd, “The PageRank citation ranking: Bringing order to
the web.,” Stanford InfoLab, 1999.
[83] Y. Malkov, A. Ponomarenko, A. Logvinov, and V. Krylov, “Approximate nearest neighbor
algorithm based on navigable small world graphs,” Inf. Syst., vol. 45, pp. 61–68, Sep. 2014, doi:
10.1016/j.is.2013.10.006.
[84] Y. A. Malkov and D. A. Yashunin, “Efficient and robust approximate nearest neighbor search using
hierarchical navigable small world graphs,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 42, no. 4,
pp. 824–836, 2018.
[85] M. Aumüller, E. Bernhardsson, and A. Faithfull, “ANN-benchmarks: A benchmarking tool for
approximate nearest neighbor algorithms,” in Proceedings of the 2017 International Conference on
Similarity Search and Applications, 2017, pp. 34–49.
[86] T. Chakraborty, A. Dalmia, A. Mukherjee, and N. Ganguly, “Metrics for Community Analysis: A
Survey,” ArXiv160403512 Phys., Apr. 2016, Accessed: Jul. 02, 2021. [Online]. Available:
http://arxiv.org/abs/1604.03512
[87] A. Hagberg, P. Swart, and D. S Chult, “Exploring network structure, dynamics, and function using
NetworkX,” Los Alamos National Lab.(LANL), Los Alamos, NM (United States), 2008.
[88] G. Rossetti, L. Milli, and R. Cazabet, “CDLIB: a python library to extract, compare and evaluate
communities from complex networks,” Appl. Netw. Sci., vol. 4, no. 1, p. 52, Dec. 2019, doi:
10.1007/s41109-019-0165-9.
[89] F. Pedregosa et al., “Scikit-learn: Machine learning in Python,” J. Mach. Learn. Res., vol. 12, pp.
2825–2830, 2011.
[90] T. Bonald, N. de Lara, Q. Lutz, and B. Charpentier, “Scikit-network: Graph Analysis in Python.,” J
Mach Learn Res, vol. 21, pp. 185–1, 2020.
[91] S. Spolaor, C. Fuchs, P. Cazzaniga, U. Kaymak, D. Besozzi, and M. S. Nobile, “Simpful: a
user-friendly Python library for fuzzy logic,” Int. J. Comput. Intell. Syst., vol. 13, no. 1, pp.
1687–1698, 2020.
[92] S. Raschka, “MLxtend: Providing machine learning and data science utilities and extensions to
Python’s scientific computing stack,” J. Open Source Softw., vol. 3, no. 24, p. 638, 2018.
[93] A. Angrish, B. Craver, and B. Starly, “‘FabSearch’: A 3D CAD Model-Based Search Engine for
Sourcing Manufacturing Services,” J. Comput. Inf. Sci. Eng., vol. 19, no. 4, p. 041006, Dec. 2019,
doi: 10.1115/1.4043211.
[94] V. J. Ferrero, “Data-Driven Environmentally Sustainable Product Design: A Shift Toward Increased
use of Sustainable Design Activities in the Early Design Phase,” Doctoral Dissertation, Oregon
State University, 2021.
[95] A. S. Deshmukh, A. G. Banerjee, S. K. Gupta, and R. D. Sriram, “Content-based assembly search:
A step towards assembly reuse,” Comput.-Aided Des., vol. 40, no. 2, pp. 244–261, Feb. 2008, doi:
10.1016/j.cad.2007.10.012.
[96] P. Wang, Y. Li, J. Zhang, and J. Yu, “An assembly retrieval approach based on shape distributions
and Earth Mover’s Distance,” Int. J. Adv. Manuf. Technol., vol. 86, no. 9–12, pp. 2635–2651, Oct.
2016, doi: 10.1007/s00170-016-8368-z.
[97] M. A. Robinson, “How design engineers spend their time: Job content and task satisfaction,” Des.
Stud., vol. 33, no. 4, pp. 391–425, Jul. 2012, doi: 10.1016/j.destud.2012.03.002.
[98] E. Pourjavad and R. V. Mayorga, “A comparative study and measuring performance of
manufacturing systems with Mamdani fuzzy inference system,” J. Intell. Manuf., vol. 30, no. 3, pp.
1085–1097, Mar. 2019, doi: 10.1007/s10845-017-1307-5.
[99] E. H. Mamdani and S. Assilian, “An experiment in linguistic synthesis with a fuzzy logic
controller,” Int. J. Man-Mach. Stud., vol. 7, no. 1, pp. 1–13, Jan. 1975, doi:
10.1016/S0020-7373(75)80002-2.
[100] X. Su and T. M. Khoshgoftaar, “A survey of collaborative filtering techniques,” Adv. Artif. Intell.,
vol. 2009, 2009, doi: 10.1155/2009/421425.
[101] J. Han, J. Pei, Y. Yin, and R. Mao, “Mining frequent patterns without candidate generation: A
frequent-pattern tree approach,” Data Min. Knowl. Discov., vol. 8, no. 1, pp. 53–87, 2004.
[102] H. Ko, P. Witherell, N. Y. Ndiaye, and Y. Lu, “Machine Learning based Continuous Knowledge
Engineering for Additive Manufacturing,” in 2019 IEEE 15th International Conference on
Automation Science and Engineering (CASE), Aug. 2019, pp. 648–654. doi:
10.1109/COASE.2019.8843316.
[103] A. Saeedi, E. Peukert, and E. Rahm, “Incremental Multi-source Entity Resolution for Knowledge
Graph Completion,” in The Semantic Web, vol. 12123, A. Harth, S. Kirrane, A.-C. Ngonga Ngomo,
H. Paulheim, A. Rula, A. L. Gentile, P. Haase, and M. Cochez, Eds. Cham: Springer International
Publishing, 2020, pp. 393–408. doi: 10.1007/978-3-030-49461-2_23.
[104] S. Fakhraei, D. Sridhar, J. Pujara, and L. Getoor, “Adaptive neighborhood graph construction for
inference in multi-relational networks,” ArXiv Prepr. ArXiv160700474, 2016.
[105] K. Teru, E. Denis, and W. Hamilton, “Inductive relation prediction by subgraph reasoning,” in
Proceedings of the 37th International Conference on Machine Learning, 2020, pp. 9448–9457.
Appendix A: Available Part Annotations

Description Name Type Classes or Text Description

component_type Single Label {‘Standard Part’, ‘Non-Standard Part’,

‘Assembly’}

features Multi Label (x3 unique {‘Holes’, ‘Flat’, ‘Cylinder’, ‘Fillet’,

features) ‘Slot’, ‘Other’, ‘Counterbore’,
‘Chamfer’, ‘Pocket’, ‘Threads’,
‘Pattern’, ‘Sweep’, ‘Freeform’, ‘Loft’}

feature_descriptions String x3 (describing Eg: ‘lengthwise semi-circular grooves’

each of the ‘features’)
Eg: ‘Two slots on the cover attached to
the motor drill’

main_process Single Label {‘Joining’, ‘Casting’, ‘Machining’,

‘Forming’, ‘Additive Manufacturing’,
‘Molding’, ‘Other’, ‘Engraving’}

material_category Single Label {‘Metal Alloys’, ‘Plastics & Polymers’,

‘Composites’, ‘Ceramics’, ‘Other’}

model_type Single Label {‘Solid Model’, ‘Surface Model’,

‘Invalid 3D Model’}

part_category Single Label {‘Home Goods’, ‘Industrial

Component’, ‘Automobile’, ‘Oil & Gas’,
‘Electronics’, ‘Electrical’,
‘Construction’, ‘Children Toy’, ‘Paper
Products’, ‘General Transportation’,
‘Cosmetics & Jewelry’, ‘Wood &
Furniture’, ‘Aerospace’, ‘Educational’,
‘Defense’, ‘Medical Device’,
‘Agriculture, Fisheries, Fishing’}

part_description String Eg: ‘hexagonal extruded part with two

concentric holes one of which is through
the part’
Eg: ‘irregular shaped block with C
shaped pocket and two cylinders
attached on one end’

precision Single Label {‘Standard Tolerance’, ‘High Tolerance’,

‘Unknown’}

quantity Single Label {‘Mass Production’, ‘One Quantity’,

‘Limited Quantity’}
Appendix B: Class-wise Mean Average Precision (mAP) Values
Table B: Comparison of HNSW and KD-Tree for retrieval using Mean Average Precision (mAP) at k = 1,
3, 5: For each class
HNSW + Exhaustive K-NN KD-Tree
Class mAP@1 mAP@3 mAP@5 mAP@1 mAP@3 mAP@5
Bearings 1 1 1 1 1 0.9918
Brackets 0.55 0.6583 0.6311 0.8 0.8083 0.8068
Bushing_Damping_Liners 0.3 0.35 0.3772 0.25 0.2833 0.3033
Bushing 0.7 0.7375 0.7421 0.525 0.5583 0.5703
Collets 0.35 0.4 0.3875 0.25 0.3 0.3
Gasket 0.95 0.95 0.9342 0.8 0.8166 0.7758
Grommets 0.8 0.8 0.8 0.75 0.75 0.75
HeadlessScrews 0.55 0.6083 0.6344 0.35 0.3416 0.3502
Hex_Head_Screws 0.85 0.875 0.8879 0.65 0.725 0.7463
Keyway_Shaft 0.5 0.5167 0.5239 0.6 0.5916 0.5902
Machine_Key 0.6 0.6167 0.6152 0.85 0.85 0.85
Nuts 1 1 1 1 1 1
O_Rings 1 1 1 1 1 1
Pipe_Fittings 0.85 0.925 0.9057 0.35 0.5791 0.6096
Pipe_Joints 0.1 0.3208 0.4232 0.15 0.3666 0.4481
Pipes 0.5 0.5583 0.5788 0.5 0.5458 0.5433
Push_Rings 0.9 0.9417 0.8761 0.9 0.9291 0.9422
Retaining_Rings 1 0.9917 0.9902 1 1 1
Rollers 1 1 1 0.9 0.9 0.9
Rotary_Shaft 0.6 0.6 0.6 0.65 0.65 0.65
Shaft_Collar 1 1 1 0.95 0.95 0.95
Slotted_Flat_Head_Screws 0.1 0.1 0.1 0 0 0.01
Socket_Head_Screws 1 1 1 1 1 1
Springs 0.8 0.7917 0.7902 0.85 0.85 0.8537
Sprockets 0.95 0.9416 0.9255 0.9 0.9083 0.8925
Thumb_Screws 0.05 0.1167 0.1592 0.05 0.0666 0.0891
Unthreaded_Flanges 1 1 1 1 1 1
Washers 0.65 0.6958 0.7078 0.3 0.3 0.3
Socket_Head_Screws 0.95 0.95 0.95 0.95 0.9333 0.9223
Appendix C: Examples of Retrieved Components for Standard Classes using HNSW+KNN Method
(Bearings, Brackets, Pipes and Springs)
C.1. Bearings: mAP@5 = 1 (Top row shows the query parts)

C.2. Brackets: mAP@5 = 0.63 (Top row shows the query parts)
C.3. Pipes: mAP@5 = 0.58 (Top row shows the query parts)
View publication stats

C.4. Springs: mAP@5 = 0.79 (Top row shows the query parts)