Data-Driven Model Categorization: Advancing Physical Systems

Analysis Through Graph Neural Networks

Andrija Grbavac

, Martin Angerbauer

, Michael Grill

and Andr

e Casal Kulzer

Research Institute for Automotive Engineering and Powertrain Systems Stuttgart (FKFS), University of Stuttgart,

Pfaffenwaldring 12, Stuttgart, Germany

Institute of Automotive Engineering (IFS), University of Stuttgart, Stuttgart, Germany

Keywords:

Graph Neural Networks, Physical System Models, Application of AI.

Abstract:

Efﬁciently categorizing physical system models is crucial for data science applications in scientiﬁc and engi-

neering realms, facilitating insightful analysis, control, and optimization. While current methods, often relying

on Convolutional Neural Networks (CNNs), effectively handle spatial dependencies in image data, they strug-

gle with intricate relationships inherent in physical system models. Our research introduces a novel approach

employing Graph Neural Networks (GNNs) to enhance categorization. GNNs excel in modeling complex

relational structures, making them apt for analyzing interconnected components within physical systems rep-

resented as graphs. Leveraging GNNs, our methodology treats entities as system components and edges as

their arrangements, effectively learning and exploiting inherent dependencies and interactions. The proposed

GNN-based approach outperforms CNN-based methods across a dataset of 55 physical system models, elim-

inating limitations observed in CNN approaches. The results underscore GNNs’ ability to discern subtle

interdependencies and capture non-local patterns, enhancing the accuracy and robustness of model categoriza-

tion in a data science framework. This research contributes to advancing model categorization, emphasizing

the application of data science for understanding and controlling complex physical systems. The innovative

use of GNNs opens new avenues for revolutionizing the categorization of intricate physical system models in

scientiﬁc and engineering domains.

1 INTRODUCTION

In the dynamic landscape of automotive engineering,

the design of powertrains has evolved exponentially,

marked by the integration of numerous sophisticated

concepts to enhance efﬁciency, performance, and sus-

tainability (Mirzadeh Phirouzabadi et al., 2020). The

complexity of modern automotive powertrains has

led to many combinations of interconnected subsys-

tems, creating an urgent need for robust modeling

techniques to help engineers understand and opti-

mize these complicated systems. Among the various

modeling approaches, physical system models have

emerged as indispensable tools, providing a compre-

hensive representation of the underlying dynamics

and interactions within physical systems (Wellstead,

1979).

These models capture the nuanced relationships

between components and their dynamic behaviors, of-

fering engineers valuable insights for analysis, con-

trol, and optimization. As the automotive industry

navigates the realms of hybridization, electriﬁcation,

and advanced control strategies (Boulanger et al.,

2011; Conway et al., 2021; Cook et al., 2002), ac-

curate and efﬁcient mastery of these models has be-

come paramount. Throughout this process, various

concepts and, therefore, physical system models are

examined, leading to an increasing number of models

in the database. This paper delves into the signiﬁ-

cance of categorizing physical system models within

the context of automotive powertrain design, empha-

sizing its pivotal role in addressing the growing intri-

cacies of modern vehicle propulsion systems.

Recognizing the limitations of conventional

methodologies, such as Convolutional Neural Net-

works (CNNs) (Grbavac et al., 2023), in effectively

analyzing physical system models, our research pro-

pounds an innovative approach. We advocate for

the application of Graph Neural Networks (GNNs),

which excels in describing, modeling, and analyz-

ing complex relational structures (Hamilton, 2020),

to augment the categorization and analysis of these

Grbavac, A., Angerbauer, M., Grill, M. and Kulzer, A.

Data-Driven Model Categorization: Advancing Physical Systems Analysis Through Graph Neural Networks.

DOI: 10.5220/0012747100003756

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 13th International Conference on Data Science, Technology and Applications (DATA 2024), pages 277-284

ISBN: 978-989-758-707-8; ISSN: 2184-285X

277

dynamic and interconnected systems. This novel per-

spective extends the current understanding of phys-

ical system models and positions GNNs as promis-

ing tools for unraveling the complexities inherent in

automotive powertrain designs. Through this explo-

ration, we aim to contribute to the ongoing discourse

on advancing engineering methodologies, fostering a

deeper understanding of the intricate dynamics shap-

ing the future of automotive propulsion systems.

The objective of this paper is to address the chal-

lenges and explore novel methods in categorizing

physical system models, a critical task with implica-

tions across various domains. In Chapter 2, we delve

into the background methods of model categorization,

emphasizing the signiﬁcance of this process and high-

lighting existing challenges faced in accurately cate-

gorizing diverse models. We introduce the concept

of graph representation as a promising approach for

modeling physical systems, setting the stage for our

subsequent exploration. Chapter 3 presents our novel

methodology, leveraging GNNs for model categoriza-

tion. We provide an overview of our approach, detail-

ing how GNNs offer a unique and effective strategy

for this task. In Chapter 4, we undertake a compre-

hensive performance evaluation, beginning with a de-

scription of our dataset and the network parameters

used. We then present our results, analyzing trends in

test accuracy achieved through our proposed method-

ology. Finally, in Chapter 5, we discuss the out-

comes and implications of our study. Our approach

not only outperforms traditional CNN methods but

also exhibits increased capacity in handling diverse

categories of models. Moreover, our method shows

promise for enabling node-level analysis and demon-

strates improved time performance compared to exist-

ing techniques. Through this structured investigation,

we aim to advance the ﬁeld of model categorization

and contribute to more effective approaches for un-

derstanding and analyzing complex physical systems.

2 BACKGROUND

2.1 Signiﬁcance of Categorizing Models

Categorizing physical system models is imperative

for harnessing the wealth of data stored within struc-

tured, hierarchical formats like Extensible Markup

Language (XML) as in this paper. These structures

contain information about the architecture of the mod-

els by linking parts together. The parts contain phys-

ical properties and, therefore, speciﬁc behavior in the

system (Wellstead, 1979). This holds signiﬁcant im-

portance in unlocking the wealth of knowledge stored

within databases, enabling engineers to leverage past

experiences and insights effectively. In engineering

and simulation domains, where the landscape is rich

with diverse models and simulations, the ability to

categorize and retrieve relevant models efﬁciently is

paramount. However, this process often proves to be

time-consuming and labor-intensive, hindering engi-

neers from fully capitalizing on the wealth of knowl-

edge accumulated from past projects.

One of the challenges engineers face is the ar-

duous task of searching for similar simulation mod-

els amidst vast databases. Without a systematic cat-

egorization framework in place, engineers must sift

through an extensive array of models, consuming

valuable time and resources. Engineers on an expert

level beneﬁt in general from their long term experi-

ence, whereas beginners rely on input to work efﬁ-

ciently and learn the physical systems fast.As signiﬁ-

cant ﬂuctuations occur and the demand for new career

entrants increases in response to employee shortages

(Akyazi et al., 2020), the imperative for implement-

ing such tools intensiﬁes. The lack of efﬁciency not

only impedes productivity but also limits the ability

to derive meaningful insights and solutions from past

projects. As a result, valuable knowledge and expe-

rience remain underutilized, preventing organizations

from maximizing the return on investment in simula-

tion and modeling efforts.

Regarding the machine learning approach, data

scientists encounter the challenge of dealing with

small numbers of physical system models that con-

tain speciﬁc concepts or architectures. Unlike other

machine learning applications where training sets are

often abundant, the availability of models containing

particular architectures or conﬁgurations may be lim-

ited in engineering domains. In some cases, data sci-

entists may only have access to one or two models

that encompass a speciﬁc architecture, posing signiﬁ-

cant challenges for training and validation purposes.

The scarcity of relevant training data compounds

the difﬁculty of developing accurate and robust cate-

gorization models, as traditional machine learning al-

gorithms as proposed in (Grbavac et al., 2023) may

struggle to generalize effectively with limited samples

(Raudys et al., 1991).

In summary, the signiﬁcance of categorizing mod-

els lies in its ability to streamline knowledge dis-

covery, enhance productivity, and facilitate informed

decision-making in engineering and simulation do-

mains. By overcoming the challenges associated with

database searchability and data scarcity, engineers can

harness the full potential of model data, driving inno-

vation and advancement in their respective ﬁelds.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

278

2.2 Existing Challenges in Model

Categorization

Despite the abundance of information encoded in

XML structures, much of it remains untapped, repre-

senting a reservoir of valuable insights waiting to be

unlocked. Efﬁciently categorizing models based on

this data becomes crucial for comprehensive model

analysis and leveraging metadata for various applica-

tions, such as search engines and predictive modeling.

Presently, there exists a prevalent approach that

utilizes CNNs to categorize these models (Grbavac

et al., 2023). However, this method encounters limita-

tions when confronted with models that exhibit mul-

tiple branches. If parts of a model, that are charac-

teristic for a speciﬁc category, are spread over multi-

ple branches, the pre-processing of the approach us-

ing CNNs, by design, struggles to capture the rela-

tionships within a model. This leads to incomplete

categorization and compromised analysis.

One alternative approach proposes enhancing the

CNN methodology by dissecting each pathway within

the model, from its inception to termination nodes.

By scrutinizing individual paths and calculating

cross-correlations along each path, it becomes possi-

ble to identify overlaps and therefore branches within

the model. However, this approach introduces com-

putational challenges, notably the time-consuming

nature of cross-correlation calculations, which exhibit

a computational complexity of O(n∗ m), where n rep-

resents the number of different paths and m denotes

their average length (Hale, 2006). Despite its poten-

tial, this method’s efﬁciency is hindered by its compu-

tational demands, necessitating further exploration of

alternative techniques for scalable model categoriza-

tion.

2.3 Introduction to Graph

Representation of Physical System

Models

As introduced in (Wellstead, 1979), physical sys-

tem models are intricate constructs that abstract real-

world systems into interconnected components, such

as mechanical, thermal, ﬂuid, magnetic, electrical ele-

ments, or equations of state, mathematical operations

and physical properties. Each component, or object,

within these models possesses distinct properties and

behaviors, representing various physical, chemical,

and mathematical elements.

Conceptually, these models can be viewed as

graphs, where nodes represent individual compo-

nents, and edges denote the connections or interac-

tions between them. As properties ﬂow between in-

terconnected components, the graph evolves dynam-

ically, reﬂecting the propagation of information and

the exchange of attributes within the system. Through

this interconnected network of nodes and edges, phys-

ical system models encapsulate the complex relation-

ships and dependencies inherent in real-world sys-

tems.

By representing physical system models as

graphs, we can leverage graph-based methodologies

to analyze and categorize these models effectively.

Graphs provide a natural abstraction for capturing

the structural and functional aspects of physical sys-

tems, enabling us to explore their intricate intercon-

nections and emergent behaviors in a systematic man-

ner (Hamilton, 2020).

3 THE NOVEL APPROACH

3.1 Graph Neural Networks in Model

Categorization

GNNs offer a powerful framework for analyzing

and categorizing graph-structured data, making them

well-suited for addressing the complexities inher-

ent in physical system models. Unlike traditional

machine learning approaches that operate on ﬁxed-

dimensional data representations, GNNs can directly

operate on graph-structured data, preserving the spa-

tial and relational information encoded within the

graph (Hamilton, 2020).

GNNs excel in learning representations of nodes

within a graph by aggregating information from their

neighboring nodes iteratively. Through a process

known as message passing, GNNs propagate informa-

tion through the graph, updating node representations

based on both local and global context. This enables

GNNs to capture the hierarchical structure and de-

pendencies present in physical system models, allow-

ing them to discern intricate patterns and relationships

that may be overlooked by conventional methods. A

more detailled overview can be found in (Hamilton,

2020).

By leveraging the rich structural information em-

bedded within graphs, the novel approach examines if

GNNs can categorize physical system models based

on their inherent properties and interconnections, and

facilitate more accurate and robust categorization out-

comes.

Data-Driven Model Categorization: Advancing Physical Systems Analysis Through Graph Neural Networks

279

3.2 Methodology Overview

The proposed methodology for categorizing physical

system models can be divided in three key steps: data

extraction and pre-processing, model building, and

training and validation:

1. Extraction of Data from XML Files and

Labelling.

The simulation model ﬁles often store much in-

formation, including parameters, setups, or ob-

ject deﬁnitions. This information has to be ﬁl-

tered. The extraction process in this methodology

involves parsing the XML ﬁles to retrieve parts

with their speciﬁc IDs and types and the links be-

tween them. A part is an element such as pipes,

compressors, actuators, or valves as depicted in

Figure 1. These extracted components and the

links between them serve as the basis for con-

structing the graph representation of the physical

system model. On top, each model is labeled with

its corresponding multiclass categories. Pairing

these two leads to a fundamental dataset to train

the neural network.

FUEL

ECU

category

Figure 1: Example of a physical system model of a com-

bustion engine including its periphery. Parts are connected

with each other via links and form a graph of physical enti-

ties. The categories of the model (framed in green) contain

a subgraph of the model (Grbavac et al., 2023).

2. Building a Graph Representation.

Using the extracted parts as nodes and the links as

connections, a graph representation of the phys-

ical system model is constructed. Each node in

the graph corresponds to a speciﬁc part of the

model, while the edges represent the connections

between these parts. The nodes can have various

features, with template IDs being utilized in this

methodology to represent object types or behav-

iors inherent in the simulation model.

3. Building a GNN Architecture.

As shown in Figure 2 the GNN architecture

(Gilmer et al., 2017; Chollet et al., 2024) consists

of several key components designed to effectively

extract features from the graph representation of

the physical system model and perform catego-

rization:

(a) Message Passing.

Message passing is performed with a spec-

iﬁed number of steps, involving information

exchange between nodes within a given n-

hop-neighborhood. This iterative process al-

lows nodes to aggregate information from their

neighboring nodes, capturing local and global

context within the graph structure.

(b) Readout.

Following message passing, a readout mecha-

nism is applied to generate graph-level repre-

sentations. This process involves partitioning

the graph into subgraphs. These then follow

padding, multi-head attention with a speciﬁed

number of attention heads, projection using se-

quential dense layers, and two-layer normaliza-

tion, followed by average pooling to obtain a

consolidated representation of the graph.

A dense layer is introduced to further reﬁne the

learned features and prepare them for catego-

rization. This layer facilitates the extraction of

higher-level features that are relevant for distin-

guishing between different categories of physi-

cal system models.

(d) Dense Output for Categories.

Finally, a dense output layer with softmax ac-

tivation is employed to produce probability

distributions over predeﬁned categories. This

layer maps the learned features to the respec-

tive categories, enabling the model to predict

the most likely category for a given physical

system model.

4. Model Evaluation and Validation.

(a) Training and Validation.

The constructed GNN architecture is trained

using a dataset comprising labeled physical

system models. During training, the model

learns to map input graph representations to

their corresponding categories through iterative

optimization of model parameters.

(b) Cross-Validation.

To assess the robustness and generalization ca-

pability of the categorization model, K-Fold

cross-validation is employed. This involves

partitioning the dataset into training sets mul-

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

280

tiple times, iteratively training and evaluating

the model on different subsets of the data.

node_features

InputLayer

input

[(None, 2)]

output

[(None, 116)]

bond_features

InputLayer

input

[(None, 2)]

output

[(None, 2)]

pair_indices

InputLayer

input

[(None, 2)]

output

[(None, 2)]

message_passing

MessagePassing

input

[(None, 49), (None, 2), (None, 2)]

output

[(None, 128)]

pair_

indicator

InputLayer

input

[(None, )]

output

[(None, )]

transformer_encoder_readout

Readout

input

[(None, 128), (None, )]

output

[(None, 128)]

dense_1

Dense

input

[(None, 128)]

output

[(None, 64)]

dense_1

Dense

input

[(None, 64)]

output

[(None, 49)]

Figure 2: Scheme of the network architecture used in this

study.

4 PERFORMANCE EVALUATION

4.1 Dataset Description

The dataset used in this study primarily consists of

system models of automotive powertrain systems,

with an overrepresented amount of internal combus-

tion engines. This emphasis on combustion engines

stems from their historical dominance in the automo-

tive industry, resulting in a larger availability of simu-

lation models for such powertrains compared to alter-

native propulsion systems.

The labeling of the dataset focuses on subsys-

tems that are characteristic for powertrain concepts.

One example are turbochargers, which serve as dis-

tinguishing factors between naturally aspirated en-

gines and turbocharged engines. The latter cate-

gory can be further subdivided into single- and two-

staged turbocharged engines, reﬂecting variations in

turbocharging conﬁgurations commonly found in au-

tomotive applications.

Furthermore, the labels in the dataset exhibit vari-

ability in size and variance, leading to differences

in complexity and detail across different categories.

This variation reﬂects the diverse nature of powertrain

conﬁgurations and allows an evaluation of the catego-

rization performance across a spectrum of complexi-

ties.

The dataset comprises 55 GT-POWER simulation

models (Gamma Technologies, 2024), each contain-

ing approximately 50 to 300 objects representing var-

ious components and subsystems of automotive pow-

ertrains. These models capture different aspects of

powertrain design and operation, ranging from basic

conﬁgurations to more intricate systems. Based on

these concepts the models are labelled by 49 cate-

gories, ranging from 1 object up to subgraphs of 143

objects.

4.2 Network Parameters

The experimental setup involves exploring a range of

hyperparameters to optimize the performance of the

categorization model. The following parameters are

varied to assess their impact on model performance:

• Message Passing Steps. The number of message

steps during message passing ranges from 1 to 16,

allowing for the exploration of different levels of

information propagation through the graph.

• Number of Attention Heads. The number of at-

tention heads in the transformer encoder readout

ranges from 1 to 15, enabling the model to capture

varying degrees of attention and focus on different

aspects of the graph representation.

• Number of Dense Units. The number of units in

the dense layer for categorization ranges from 32

to 1024, inﬂuencing the complexity and expres-

siveness of the learned feature representations.

Additionally, to ensure robust evaluation and vali-

dation of the categorization model, the dataset is split

into 10 folds using a K-Fold cross-validation strategy.

A binary cross-entropy loss function is chosen and op-

timized by an Adaptive Momentum (Adam) optimizer

as it acchieves great results with graph classiﬁcation

(Gilmer et al., 2017).

4.3 Results

The categorization model demonstrates excellent per-

formance overall, achieving an accuracy of over 90%

on the test dataset. This high level of accuracy high-

lights the model’s ability to effectively categorize

physical system models, even when faced with com-

plex categories and variations in model characteris-

tics.

Figure 3 shows the average test accuracy for all

49 classes. These are shown for the parameter com-

bination with the best overall results. Regarding the

Data-Driven Model Categorization: Advancing Physical Systems Analysis Through Graph Neural Networks

281

categorization ability for each class, the best results

are achieved with 4 message steps, 10 attention heads,

and 32 dense units. This conﬁguration achieves very

good accuracy for each class, with the exception of

class no. 4 (Single Staged Turbocharged Engine),

which exhibits slightly lower accuracy. However, all

other classes achieve at least 83% accuracy, demon-

strating the model’s strong categorization capabilities

across diverse categories.

Figure 3: Average accuracy of the 49 categories after a

training with 4 message steps, 10 attention heads and 32

dense units over 10 folds. With an overall accuracy of 98%

this parameter combination forms the best overall results.

Trends of the Test Accuracy.

Several notable trends emerge from the experimental

results, providing insights into the impact of hyperpa-

rameters on categorization performance:

• When using 32 neurons in the categorizing dense

layer, the median of the test accuracy remains con-

sistently high regardless of the number of mes-

sage steps and attention heads (see Figure 4a-c).

However, as shown in (see Figure 4d-f and 4g-i),

with an increasing number of neurons in the dense

layer, the median tends to decrease as the number

of message steps increases.

• As shown in see Figure 4a-c, the number of atten-

tion heads shows a small inﬂuence on the mean

accuracy when the network has 32 dense units

in the categorization layer, with hardly better re-

sults observed with a higher number of attention

heads. However, this trend is not consistent when

using a higher number of dense units, indicating a

more complex relationship between these param-

eters (see Figure 4g-i).

• Overall, the number of attention heads has no

signiﬁcant inﬂuence on accuracy, suggesting that

other factors play a more dominant role in deter-

mining categorization performance.

• However, the variance of accuracy over the k folds

increases with a higher number of dense units in

the categorizing layers ( Figure 4a, d and g) and a

higher number of message steps Figure 4c, f and

i), indicating increased sensitivity to variations in

these hyperparameters.

5 DISCUSSION

The results obtained from the experimental evaluation

of the novel GNN based approach for categorizing

physical system models reveal signiﬁcant improve-

ments over the traditional CNN approach (Grbavac

et al., 2023). Several key points highlight the supe-

riority of the GNNs approach and suggest potential

areas for further enhancement and analysis.

5.1 Superior Performance over CNN

Approach

The performance of the novel GNN-based approach

signiﬁcantly surpasses that of the CNN approach. The

CNN approach, which extracts paths from start to

end nodes and categorizes them individually, is in-

herently limited in its ability to handle categories that

span multiple branches (see Figure 5). In contrast,

the GNN approach demonstrates superior adaptabil-

ity and robustness in categorizing models with com-

plex structural relationships that extend across multi-

ple branches. This highlights the inherent advantages

of the GNN approach in handling graph data with in-

tricate interconnections.

5.2 Increased Capacity for Handling

Categories

The GNN approach exhibits a much higher capac-

ity for handling categories compared to the CNN ap-

proach. While the categories in the CNNs paper are

limited to a maximum of 15 objects, the novel ap-

proach extends this limitation signiﬁcantly, accom-

modating categories consisting of up to 143 objects.

This expansion not only increases the breadth of cat-

egories that can be considered but also enhances the

complexity of these categories, allowing for more nu-

anced and detailed categorization of physical system

models.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

282

(a) (b) (c)

(d) (e) (f)

(g) (h) (i)

Figure 4: Investigation of parameter variations on the test accuracy of the GNN model. The boxplots represent the K-Fold

distribution for different parameter combinations. The box plot highlights the median, quartiles, and outliers, providing

insights into the variability and central tendencies of the test accuracy.

start-node

end-node

Model paths:

1  2  3  4  7  10  13  14

1  2  3  5  8  11  13  14

1  2  3  6  9  12  13  14

object

connection

Categor X

Figure 5: Example of a branched graph and the extraction of

the model paths based on the CNN approach. To categorize

all elements of the category X, the paths have to be catego-

rized based on a characteristic length in each path. Other-

wise an intersection has to be calculated via cross-entropy

(Grbavac et al., 2023).

5.3 Potential for Node-Level Analysis

To further analyze the differences in accuracies result-

ing from different parameter combinations, the novel

approach can be extended to enable node-level anal-

ysis. By examining embeddings on a node level and

comparing them with node-level labeling, the perfor-

mance of the GNN can be assessed in greater detail.

This extension would provide insights into how effec-

tively the GNN model captures and utilizes informa-

tion at the level of individual nodes, shedding light on

its performance beyond the graph level.

5.4 Improved Time Performance

In addition to its superior categorization performance,

the novel GNN approach also outperforms the CNN

approach in terms of time efﬁciency. The CNN ap-

proach’s preprocessing phase is signiﬁcantly more

time-consuming due to its large recursive search

Data-Driven Model Categorization: Advancing Physical Systems Analysis Through Graph Neural Networks

283

through the graph and the training process involving

a substantial number of extracted paths. In contrast,

the GNN approach achieves comparable categoriza-

tion accuracy with approximately ten times faster pro-

cessing times, making it a more practical and efﬁcient

solution for real-world applications.

6 CONCLUSIONS

Tthe present study has demonstrates the effective-

ness of a novel GNN based approach for categoriz-

ing physical system models, particularly focusing on

automotive powertrain systems. Through rigorous ex-

perimentation and analysis, several key ﬁndings have

emerged, highlighting the signiﬁcant advancements

achieved by the proposed methodology.

Firstly, the GNN approach shows superior perfor-

mance compared to traditional CNN methods, par-

ticularly in handling complex graph structures with

branched pathways. While the CNN approach strug-

gles with categorizing models that span multiple

branches, the GNN approach, leveraging message

passing, exhibits remarkable adaptability and robust-

ness in capturing the intricate interconnections within

the graph.

Furthermore, the GNN approach indicates an in-

creased capacity for handling a wider range of cat-

egories, including those with higher complexity and

variance. By extending the limitations imposed by

previous CNN-based methods, the novel approach en-

ables more reﬁned categorization of physical system

models, thereby enhancing the depth of analysis and

insights derived from the categorization process.

Moreover, the potential for node-level analysis

presents exciting opportunities for further reﬁnement

and optimization of the GNN approach. By examin-

ing embeddings at the node-level it can provide valu-

able insights into its effectiveness in capturing infor-

mation at a granular level.

Finally, the improved time efﬁciency of the GNN

approach gives a signiﬁcant practical advantage over

traditional CNN methods. With processing times

approximately ten times faster than CNN-based ap-

proaches, the GNN approach offers a more efﬁcient

and scalable solution for real-world applications.

In summary, the ﬁndings from this study under-

score the promising potential of GNN-based method-

ologies for advancing the ﬁeld of system model cat-

egorization. As we continue to reﬁne and optimize

these approaches, we can expect further advance-

ments in our ability to analyze and understand com-

plex physical systems, ultimately driving innovation

and progress in engineering and simulation research.

REFERENCES

Akyazi, T., Alvarez, I., Alberdi, E., Oyarbide-Zubillaga, A.,

Goti, A., and Bayon, F. (2020). Skills needs of the

civil engineering sector in the european union coun-

tries: Current situation and future trends. Applied Sci-

ences, 10(20):7226.

Boulanger, A. G., Chu, A. C., Maxx, S., and Waltz, D. L.

(2011). Vehicle electriﬁcation: Status and issues. Pro-

ceedings of the IEEE, 99(6):1116–1138.

Chollet, F. et al. (2024). Keras. https://keras.io.

Conway, G., Joshi, A., Leach, F., Garc

ıa, A., and Senecal,

P. K. (2021). A review of current and future power-

train technologies and trends in 2020. Transportation

Engineering, 5:100080.

Cook, J., Sun, J., and Grizzle, J. (2002). Opportunities in

automotive powertrain control applications. In Pro-

ceedings of the International Conference on Control

Applications, volume 1, pages xlii–xlli vol.1.

Gamma Technologies (2024). GT-SUITE.

Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and

Dahl, G. E. (2017). Neural message passing for quan-

tum chemistry.

Grbavac, A., Angerbauer, M., Grill, M., Itzen, D., Miloje-

vic, S., Hagenbucher, T., and Kulzer, A. (2023). Cat-

egorizing simulation models using convolutional neu-

ral networks. Technical report, SAE Technical Paper.

Hale, D. (2006). An efﬁcient method for computing local

cross-correlations of multi-dimensional signals. CWP

Report, 656:282.

Hamilton, W. L. (2020). Graph representation learning.

Morgan & Claypool Publishers.

Mirzadeh Phirouzabadi, A., Savage, D., Blackmore, K., and

Juniper, J. (2020). The evolution of dynamic interac-

tions between the knowledge development of power-

train systems. Transport Policy, 93:1–16.

Raudys, S. J., Jain, A. K., et al. (1991). Small sample

size effects in statistical pattern recognition: Recom-

mendations for practitioners. IEEE Transactions on

pattern analysis and machine intelligence, 13(3):252–

264.

Wellstead, P. E. (1979). Introduction to physical system

modelling, volume 4. Academic Press London.

ACRONYMS

Adam Adaptive Momentum.

CNN Convolutional Neural Network.

GNN Graph Neural Network.

XML Extensible Markup Language.

DATA 2024 - 13th International Conference on Data Science, Technology and Applications

284