Enhancing Multi-objective Optimisation

Through Machine Learning-Supported

Multiphysics Simulation

Diego Botache1(B) , Jens Decke1 , Winfried Ripken2 , Abhinay Dornipati3 ,

Franz Götz-Hahn1 , Mohamed Ayeb3 , and Bernhard Sick1
1
Intelligent Embedded Systems, University of Kassel, Hessen, Germany
{diego.botache,jens.decke,franz.goetz.hahn,bsick}@uni-kassel.de
2
Merantix Momentum GmbH, Berlin, Germany
3
Vehicle Systems and Fundamentals of Electrical Engineering, University of Kassel,
Hessen, Germany
{abhinay.dornipati,ayeb}@uni-kassel.de
https://www.uni-kassel.de/eecs/ies, https://www.merantix-momentum.com/,
https://www.uni-kassel.de/eecs/fsg

Abstract. This paper presents a methodological framework for train-

ing, self-optimising, and self-organising surrogate models to approximate
and speed up multiobjective optimisation of technical systems based on
multiphysics simulations. At the hand of two real-world datasets, we illus-
trate that surrogate models can be trained on relatively small amounts
of data to approximate the underlying simulations accurately. Includ-
ing explainable AI techniques allow for highlighting feature relevancy or
dependencies and supporting the possible extension of the used datasets.
One of the datasets was created for this paper and is made publicly
available for the broader scientific community. Extensive experiments
combine four machine learning and deep learning algorithms with an
evolutionary optimisation algorithm. The performance of the combined
training and optimisation pipeline is evaluated by verifying the generated
Pareto-optimal results using the ground truth simulations. The results
from our pipeline and a comprehensive evaluation strategy show the
potential for efficiently acquiring solution candidates in multiobjective
optimisation tasks by reducing the number of simulations and conserv-
ing a higher prediction accuracy, i.e., with a MAPE score under 5% for
one of the presented use cases.

Keywords: Electric Motors · Multiobjective Optimisation ·

Surrogate-Modelling · Deep-Learning · Explainable Artificial
Intelligence

⃝c The Author(s), under exclusive license to Springer Nature Switzerland AG 2024

A. Bifet et al. (Eds.): ECML PKDD 2024, LNAI 14950, pp. 297–312, 2024.
https://doi.org/10.1007/978-3-031-70381-2_19
298 D. Botache et al.

1 Introduction
Multiphysics and multiscale simulations have become crucial for the computa-
tional modelling and analysis of multiple interacting physical phenomena in tech-
nical systems. These phenomena include mechanics, fluid dynamics, heat trans-
fer, and electromagnetics for a wide variety of applications, such as aerospace
engineering, biomedical engineering, and materials science, to name a few. Incor-
porating multiple physical phenomena into simulations is a powerful tool for
engineers, enabling them to investigate various design alternatives and param-
eters and enhance the depth of their decision-making during design processes.
Typically, this approach involves considering competing objectives simultane-
ously within multiobjective optimization tasks, thereby ensuring that the final
design solutions strike an optimal balance across diverse performance criteria.
Acquiring optimal solutions presents a significant challenge due to the com-
plex nature of numerical models and potential nonlinear dependencies among
design parameters. Moreover, constrained solution spaces yield scarce feasible
solutions and require the inclusion of advanced domain knowledge of the under-
lying physical problem. To address these issues, we propose using surrogate mod-
els, i.e. data-driven algorithms, as an alternative to running computationally
expensive multiphysics simulations in combination with advanced optimisation
techniques i.e. evolutionary algorithms. We demonstrate how surrogate mod-
els in multiphysics simulations reduce computational burdens, accelerating the
design process and enabling broader exploration of design options.

Fig. 1. Proposed strategy for training and self-optimising surrogate models using
machine learning and deep learning techniques to tackle multiobjective optimisation
problems in complex multiphysics simulations.

As shown in Fig. 1, our strategy can be summarised as follows: First, we iden-

tify the multiphysics system input and output space followed by a small data
generation procedure. Next, we apply our pipeline for the training and evalua-
tion of machine learning (ML) models, which are then used for the optimisation
task of the input parameters of the multiphysics system. Finally, we validate the
optimisation results by comparing the outputs of the ML models against the sim-
ulated values of selected solution candidates of the multiobjective optimisation
problem.
The main contributions of our paper can be summarised as follows:
 MO-Optimisation Through ML-Supported Multiphysics Simulation 299

1. A real-world tabular dataset using multiphysics simulations, which is made

publicly available
2. We streamline the efficient acquisition of solution candidates in multiobjec-
tive optimisation tasks using multiple surrogate models in combination with
evolutionary algorithms
3. We validate the acquired results and derive insights about the relevance and
dependencies of features in our real-world examples by using explainable AI
techniques

The remainder of this article is structured as follows: Sect. 2 summarises the

related work, followed by a detailed description of our combined training and
optimisation pipeline in Sect. 3, Sect. 4 describes the underlying physical prob-
lems of the two real-world tabular datasets used in our approach. The results of
our evaluation strategy for training and optimisation with an examination of the
acquired results with explainable AI methods are presented in Sect. 5. Finally,
Sect. 6 summarises the main conclusions and open issues for future work.

2 Related Work

Coupling multiphysics simulations may compromise stability, accuracy, and

robustness [13,17]. Utilizing machine learning (ML) and deep learning tech-
niques to predict efficiently and approximate multiphysics processes typically
requires large amounts of data for training and feature extraction [4]. While
individual approaches to substituting simulations with ML techniques have been
proposed, there is an absence of a general strategy for training, hyperparameter
tuning, evaluating and comparing diverse data-driven models in multiphysics
simulations. Evaluations are often customised for specific applications, and it is
unclear whether the approach can be generalized to other problems, especially
in cases with sparse data.

2.1 Surrogate Modelling

When dealing with tabular data, employing boosted trees or ensemble strategies
for meta-modelling is a reliable benchmark for execution speed, computational
efficiency, and accuracy [9]. In our research, we depict the multiphysics problem
in tabular format, enabling direct training of surrogates for predicting individual
output values of the associated physical system. This common approach allows
hyperparameter tuning of each surrogate and gives insight into the models’ per-
formance for each system output. Besides fixed tabular or parameterised physical
problems, integrating deep learning techniques with numerical simulation tools
has shown the potential to reduce computational burden using more complex
data. These techniques can predict multiple outputs of a physical system and
extract in-depth features of the data. For example, convolutional auto-encoders
have been used to model scalar transport equations coupled to Navier-Stokes
equations [10]. Further application strategies in fluid flow applications [1,18],
300 D. Botache et al.

heat transfer [14,25] and electromagnetics [20] show the potential for accurate
modelling of multiphysics problems but also on predicting problem specific key
performance indicators. Still, most of these approaches lack the explainability
and interpretability of the trained or extracted features and require high amounts
of data. Physics-informed neural networks [16,23,25] have been proposed as a
powerful tool to obtain solutions for multiphysics simulations without access to
ground-truth data. However, these models require substantial knowledge of the
underlying system, making them less accessible to non-experts. Additionally,
they can be computationally expensive and time-consuming to train, needing
significant resources to converge to accurate solutions.

2.2 Multiphysics Optimisation and Data Extension

In many cases, the goal consists of optimising the design parameters within
these technical systems, considering multiobjective criteria. Multiobjective
optimisation requires finding trade-offs between potentially conflicting objec-
tives [24], which quickly becomes a challenging task for automated optimisa-
tion algorithms. A recent study investigated this problem through the lens of
diversity and showed that their approach increases diversity without sacrificing
global performance [22]. In our work, we shift the decision to consider the dif-
ferent objectives of the user by presenting a range of Pareto-fronts of optimised
designs. We additionally encourage diversity in our optimised results by carefully
choosing combinations of surrogate models and optimisation algorithms.
Limited data also poses significant challenges in multiphysics optimisation,
and only a few works highlight the challenges in specific use cases. An investi-
gation integrates topology optimisation and generative models (e.g., generative
adversarial networks) in a framework allowing the exploration of new design
options, thus generating many designs starting from limited previous design
data [19]. In a preliminary investigation of one of the use cases, novelty and
anomaly detection algorithms were used in design optimisation tasks. The study
found that these algorithms are effective in exploring the design space, but they
have limitations when it comes to exploitation [7]. Thus, Deep active design
optimization was introduced to address this disadvantage, combining deep active
learning and design optimization [6].

3 Machine Learning Supported Optimisation Strategy

In this section, we present our pipeline for applying, explaining and evaluat-
ing ML surrogate models coupled with multiobjective optimisation strategies of
technical systems. It streamlines the evaluation and experimentation process and
presents a general validation and performance evaluation procedure. Our app-
roach, depicted in Fig. 2, consists of three main blocks, which will be explained
separately in the following subsections.
 MO-Optimisation Through ML-Supported Multiphysics Simulation 301

Fig. 2. Ilustration of our proposed Pipeline for self-optimising surrogate models, which
can be seamlessly integrated into the optimisation process of multiphysics problems and
comprises three main blocks: data acquisition, surrogate model training, and multiob-
jective optimisation.

3.1 Data Acquisition

The first block (1), shown in Fig. 2, consists of the data acquisition module,
which, in the initial phase, is a manual process. Here, a small database of sim-
ulation results from available numerical models is gathered, functioning as a
starting point of the optimisation pipeline. This step is crucial for the success
of the optimisation process and poses specific requirements and considerations.
On the one hand, the initial database must cover many feasible input parameter
combinations to train the surrogate models efficiently. On the other hand, the
generation process itself relies on slow and complex numerical models, implicitly
restricting the size of the database. Naturally, the quality of this first selection of
data impacts the accuracy and reliability of the subsequently trained machine-
learning models and the outcome of the optimisation process as a whole. In
the context of the outlined use cases, the supplied tabular data comprises sin-
gular simulation instances. It can be extended to non-parametric or tabular
approaches like pictures or videos, but this is not the scope of this paper. Con-
sequently, instances characterized by incomplete simulation results or missing
data will be systematically filtered from the database. In addition, depending
on the type of surrogate model, some data pre-processing is essential and can be
defined by the pipeline user.

3.2 Surrogate Models

The extreme gradient boosting (XGB) [2] serve as our baseline surrogate model
and we train single models for the prediction of individual output values. In
contrast, deep learning techniques are used for predicting multiple target values,
302 D. Botache et al.

therefore they can consider relationships in the output space. The training of
the surrogate models is shown in Fig. 2 block (2). We employ hyperparameter
optimisation strategies to ensure robust training of the models. Since we assume
that the databases are limited and usually distributed in a specific range of
parameter values it is still possible to generate solution candidates outside that
range at the end of the proposed pipeline, but this could affect the prediction
performance of the models. Therefore, to identify possible unexplored areas that
should be generated in the first block, Fig. 2 (1), it is crucial to understand the
dependencies present in the data. To achieve this, our approach incorporates
explainable artificial intelligence (xAI) techniques that provide useful insights
into the data will be explained in detail in the following.

3.3 Interpretable Surrogate Modelling (xAI Module)

By analysing the results obtained from selected xAI methods, we can identify
features or parameters in the data that highly influence the outputs of the phys-
ical systems. Additional methods can also provide information about the dis-
tribution of the parameter values, showing areas of the data that are not fully
explored, and therefore extend the database accordingly. We used feature rele-
vances from the XGB surrogates to understand the impact of input features on
target values. Furthermore, we extended our approach using partial dependence
plots [26] to identify whether the relationship between relevant features and tar-
get values is linear, non-linear, monotonic, or more complex. These plots are a
powerful technique to show the marginal effects of one or two variables on the
predicted outcome of a machine learning model while holding all other variables
constant. Additional xAI techniques can be included in the pipeline but will not
be further considered in this study.

3.4 Multiobjective Optimisation and Validation

The third block (3) of our approach leverages the pre-trained machine learning
models to perform the multiobjective optimisation task, including evolutionary
algorithms. The optimisation algorithm used in the following use cases is a non-
dominated sorting genetic algorithm (NSGA). Genetic algorithms have proven
to be a robust and reliable design optimisation method [3]. Therefore, we inte-
grated the NSGA-2 into our optimisation pipeline [5]. The results provided by
the evolutionary algorithm will be evaluated using Pareto frontiers, and the final
selection of solution candidates will correspond to the points at the Pareto front.
A validation step is required to validate the prediction performance of the surro-
gate models within the selected solution candidates. After the validation step, we
propose using performance indicators [12,15] to compare the simulated Pareto
frontiers from multiple experiments with different surrogate models.

4 Experimental Design
In the following use cases, we explain the optimisation task and the correspond-
ing target values to consider when training the surrogate models. We train a
 MO-Optimisation Through ML-Supported Multiphysics Simulation 303

Table 1. Dimensions, parameters, con-

straints, and performance characteristics

Dimension or Parameter Value

Stator outer ! 230 [mm]
Rotor outer ! 152 [mm]
Airgap Length 1 [mm]
Magnet Remanence at 120 ℃ 1.18 [T]
Constraints Value
Maximum Current 636.34 [A]
Maximum Voltage 270 [V]
Performance Characteristics Value
JM @1000 rpm 280.5 [Nm]
JΦ @1000 rpm 3.76 [kW]
Jm 2.78 [kg]
Fig. 3. Parameterised 2D rotor segment
of the baseline machine with black fields
representing the magnet components and
grey areas for the air cavities. (Color
figure online)

single XGB Regressor for each target value considering the mean squared error
(MSE) and optimise each model’s hyperparameters independently using a com-
bined cross-validation and Bayesian optimisation strategy. The XGB as a base-
line provides valuable insights into the underlying problem and data depen-
dencies. Further, we employ ensemble strategies combining multiple scikit-learn
regressors at the decision level [11,21]. First, we conduct a random search to iden-
tify promising regressors. Next, we form an ensemble of these regressors using
a weighted average, with weights optimized via gradient-based methods. The
ensemble is trained with cross-validation, and the best one is selected based on
validation scores. The final chosen ensemble is trained on the entire training
dataset. This approach balances exploring the hyperparameter space and exploit-
ing the most promising models. Finally, we use two deep-learning methods, MLP
and CNN, for the regression task. We use both models to estimate all target
values, and we tune the hyperparameters of each model using a combined cross-
validation and Bayesian optimisation strategy.

4.1 Use Case 1: Motor Dataset

The first use case consists in optimising the performance of a topology of an elec-
tric motor. The objectives of the optimisation problem correspond to maximising
the mean Torque JM while minimising the Total Loss JΦ and the Magnet
Mass Jm . Therefore, the optimisation objectives are conflicting in nature. The
baseline machine for optimization is described by its dimensions, parameters,
constraints, and performance characteristics in Table 1. For this work, only the
rotor topology of the baseline machine—a 60-slot 10-pole internal permanent
magnet synchronous machine—is parameterised with 15 geometric parameters.
These parameters govern the size and position of the magnets and air cavities,
304 D. Botache et al.

which significantly impact all the target values.1 . As shown in Fig. 3, two lat-
eral magnets are embedded into the rotor iron lamination. In addition, two air
cavities are designed in the rotor topology to guide the magnetic flux from the
magnets into the air gap at the stator-rotor interface.
Efficient sampling of high-dimensional spaces is crucial to deliver enough
information about the design space and, thus, train ML surrogate models. This
is usually constrained by the computational resources with the numerical mod-
els, and exploring every possible combination of parameters becomes unfeasible.
In this case, we apply Latin Hypercube Design (LHD) sampling, which limits the
number of simulations and enables systematic and efficient exploration across the
parameter space of the motor topology. The COMSOL 6.0 uncertainty quantifi-
cation module generates 691 sample points based on this algorithm, i.e., motor
design variations. The training and test datasets consist of 552 and 139 designs,
respectively.

4.2 Use Case 2: U-Bend Dataset

The second use case, from fluid dynamics, addresses energy dissipation caused
by flow deflections in technical systems. Typical applications include large-scale
piping grids or cooling channels of gas turbine blades. Figure 4 depicts the U-
bend, which is described by 28 design parameters consisting of six boundary
points (green) and 16 curve parameters (red), offering considerable flexibil-
ity in its design and potential for optimisation. Two target values—Pressure
Loss Jp and Cooling Power JT —are evaluated using computational fluid
dynamics. The Navier-Stokes equations determine the velocities of the fluid and
the pressure. In addition, the energy equation is used to consider the heat con-
duction and the convective heat transfer between fluid and solid using a multi-
physics solver. Target JT is inversely defined, meaning a lower value signifies
higher cooling power. However, these two objectives inherently conflict, as no sin-
gle design can optimise both simultaneously, presenting a classic multi-objective
optimization challenge.
The dataset is publicly available [8], and the inclined reader is referred to the
relevant paper for a detailed description. We utilize a subset of the data, focusing
solely on the parameter representation of the dataset to maintain comparability.
In contrast to the Motor case, the authors used identical and independent ran-
dom sampling to generate the dataset. The training set comprises 800 designs,
and the test set includes 200.

1
Please refer to https://github.com/dbotache/ for a detailed description of the motor
dataset use case.
 MO-Optimisation Through ML-Supported Multiphysics Simulation 305

Fig. 4. Parameterised geometry with boundary points (green) and curve parameters
(red). The Figure is adapted and rotated from [7] (Color figure online)

5 Results

We illustrate the versatility of our pipeline using the presented use cases and
structure our results as follows: Subsect. 5.1 showcases the performance of the
involved ML-surrogate models. Subsection 5.2 provides valuable insights into the
regression task’s outcomes. Results obtained with an evolutionary optimisation
strategy are presented in Subsect. 5.3. Moreover, in Subsect. 5.4, we validate the
acquired solution candidates by comparing the predictions of the surrogate mod-
els against the numerical simulations and finally present a performance criterion
for selecting the final results.

5.1 Prediction Performance of Surrogate Models

Surrogate models are evaluated using the mean absolute percentage error
(MAPE) on the test dataset (20%). This metric allows evaluation and compar-
ison of the models’ performance regardless of the target’s scale, as seen in the
top half of Fig. 5. Here, MAPE values below 1.9% for the Torque (Jm ), Total
Loss (JΦ ) and Magnet Mass (Jm ) indicate good prediction performance on
the Motor dataset. The U-Bend dataset poses a more difficult problem, evident
by the significantly higher MAPE scores using any surrogate model for predicting
the target values Pressure Loss (JP ) and Cooling Power (JT ). The bottom
part of Fig. 5 shows notable residual values for specific data points, particularly
in predicting JT for the U-Bend dataset. These outliers likely contribute to the
elevated MAPE values observed for the same target variable, given the sensitivity
of this metric to extreme values.
Another potential reason for the higher prediction error in the U-Bend case
is the greater dimensionality of the input space, which has 13 dimensions more
306 D. Botache et al.

Fig. 5. Assessment of surrogate model performance using MAPE scores and residual
analysis on the test sets for each use case.

than the Motor case with 15 design parameters. MSE loss for training and hyper-
parameter optimisation leads to a severe penalty for outliers. Consequently, the
optimisation process becomes more challenging in regions featuring low target
values, as they have less importance when squared. Moreover, the U-Bend case
features an entirely turbulent flow, where minor parameter adjustments can
trigger vortex formation and introduce discontinuities in the objective function.
The cases differ fundamentally in their physics: the Maxwell equation is a first-
order linear partial differential equation, whereas the Navier-Stokes equation is a
second-order nonlinear system with five solution variables, potentially account-
ing for the variance in prediction accuracy across our two use cases.

5.2 Identifying Critical Features and Relevant Dependencies

We start considering in Fig. 6 the feature importances of each XGB model trained
independently to predict a single output value in each of the presented use cases.
In the Motor case, we can observe notable feature importances for the design
parameters wi mag1, wi mag2 and hi mag1 concerning the output value Mag-
net Mass (Jm ). It was anticipated that these parameters would strongly cor-
relate with the magnet’s mass since it is a function of its size. Regarding the
output value Total Loss (JΦ ) the high importance value for the parameter
wi ac1 can be explained by the air cavities influence for the magnetic flux from
the magnets to the air gap, which could affect the overall loss and efficiency.
Finally, the most relevant feature for the target Torque (JM ) is given by the
parameter hi mag1. This correlates with the knowledge that the magnet com-
ponents influence the intensity of the magnetic flux and, therefore, the torque.
The partial dependency plots presented in Fig. 7 underscore the identified cor-
relations using the feature importances. The curves show mainly linear depen-
dencies for the Motor case with a high influence of the width parameters of
the magnets and a low influence of the air cavity parameters on the Mag-
net Mass Jm . Conversely, air cavity-related parameters primarily influence the
 MO-Optimisation Through ML-Supported Multiphysics Simulation 307

Fig. 6. Analysis of feature importance. The plots show the stacked bars of feature
important values obtained from the XGBoost models for each target value and use
case.

Total Loss JΦ and we can notice an extremely low influence of air cavity param-
eters for the target value Jm .

Fig. 7. Partial dependency plots for selected features and target values for the Motor
and the U-Bend datasets.

In the U-Bend case, the parameter DoY is crucial for Pressure Loss (JP ),
influencing the channel width at the outer layer’s deflection point. This is
expected, since altering the channel’s width affects the local Reynolds number,
i.e. the turbulence and vortex formation that contribute to JP . Although EoX
also impacts JP , its effect is less pronounced. Outer curve parameters notably
affect JP , as optimal flow alignment can prevent vortex formation, whereas
too small parameters may cause excessive deflection and stall. For Cooling
Power JT , parameters EoX, CiX, and EiX are pivotal with the highest impor-
tance values. These parameters are especially relevant to the inner layer and
directly impact the heated surface area, where a thinner surface lowers heat
conduction resistance. Given the high Reynolds numbers, the dominant heat
transfer mechanism from the solid to the fluid is convective, making a thin solid
layer beneficial. Narrowing the channel—mainly managed by outer parameters—
enhances cooling performance by promoting efficient flow contact.
308 D. Botache et al.

5.3 Evaluation of Solution Candidates in the Multiobjective

Optimisation Task
The optimisation results acquired with the four surrogate models in combination
with the evolutionary algorithm are presented in Fig. 8. With our presented EA
optimisation strategy, we can evaluate around eight million points using the sur-
rogate models instead of the numerical simulations. The red dots represent the
data points in the original database used to train and test the surrogate models.
The coloured (Motor case left) and blue (U-Bend case right) dots correspond
to the predictions of each surrogate model for the final population (2.000 sam-
ples) of points present in the last iteration of the evolutionary algorithm. The
colour of the points (left) represents the value of the Magnet Mass Jm . The
Pareto frontiers with dark red points are based on the model’s prediction. In the
Motor case (left), we can notice that the frontiers given by the prediction from
the MLP and CNN surrogates are far from the original database. In contrast,
the results with the XGB and Ensemble surrogate models deliver solution can-
didates which only outperform the database in regions of higher torque values
(JM > 250).

Fig. 8. Optimisation Results for the Motor Case (left) and the U-Bend dataset (right)
with EA in a combination of four different surrogate models

In the U-Bend case, in Fig. 8 (right), evolutionary optimization produces

Pareto frontiers superior to those in the initial database for each surrogate
model, shown as dark red dots. However, a critical observation is that three
 MO-Optimisation Through ML-Supported Multiphysics Simulation 309

of the four models predict physically unrealistic negative pressure loss values,
suggesting a lack of understanding of the physical principles involved. Recalcu-
lating the design candidates from these frontiers will further assess the models’
performance.

5.4 Validation of Solution Candidates in the Multiobjective

Optimisation
In the last stage of our evaluation strategy, we compare the predictions of the
surrogate models, i.e. the selected solutions at the Pareto frontiers, against the
true values using the original physical numerical models. The first indicator
for the surrogate models’ quality is reflected in Table 2 through the simulation
rate. A higher simulation rate reveals the models’ ability to capture parameter
interrelationships and generate many plausible solutions.

Fig. 9. MAPE of the final validation of the Motor and U-Bend Use Cases

Fig. 10. Pareto frontiers of the final validation of the Motor and U-Bend Use Cases

In the Motor case, we achieve a 100% simulation rate with the ensemble
model, and due to the complexity of the U-Bend case, only up to 78% of the
solution candidates provided likewise with the ensemble were plausible. Next, in
Fig. 9, we validate the solution candidates using the MAPE score. Higher MAPE
scores for the U-Bend case reflect higher complexity in the design space and more
challenging physical principles. However, the deviation of model predictions in
310 D. Botache et al.

areas of interest in the optimisation task is mainly affected by the lower presence
of samples in the training database. If the simulation rate is extremely low, an
additional training cycle of the surrogate models using the simulated data points,
i.e., plausible samples from the Pareto frontiers, should be considered. These
additional data for training can provide additional information about regions of
lower density and closer to the Pareto frontiers to the surrogate models.

Table 2. Performance indicators of validated solutions candidates concerning the orig-

inal database (DB) for both use cases using Generational Distance (GD), GD-Plus
(GD+) [15] and the Hypervolume Indicator (HV) [12].

Motor Case U-Bend Case

Sim. Rate GD GD+ HV↑ Sim. Rate GD GD+ HV↑
Database (DB) – – – 1.753 – – – 0.973
DB + XGB 98.91% 0.059 0.012 1.766 69.2% 0.063 0.033 1.077
DB + Ens. 100% 0.048 0.018 1.795 78.3% 0.062 0.009 1.112
DB + MLP 80.43% 0.038 0.010 1.774 20.9% 0.111 0.078 1.091
DB + CNN 91.81% 0.041 0.012 1.770 53.6% 0.128 0.064 1.157

The numerical validation and final results are shown in Fig. 10. The optimi-
sation task was successful in both cases. In the Motor case, we appreciate better
solutions candidates at higher Torque values as appreciated in the predictions in
Subsect. 5.3. Particularly in the U-Bend case, the simulation results are closer
to the database and reflect higher deviations from the predicted values in Fig. 8
on the right. However, the obtained solutions candidates outperformed the orig-
inal database Pareto frontier in all regions, concerning the two objective values
JT and JP . These results emphasise the versatility of our strategy and set the
focus on the quality of the final solution candidates, which do not depend on the
prediction accuracy of the surrogate models.
In our experiments, four surrogate strategies were compared using Pareto
frontiers, but combining more surrogates and optimization strategies could
make qualitative comparisons unmanageable. Using performance indicators as
shown in Table 2 allows a quantitative comparison of the validated Pareto fron-
tiers and the filtering of only the most significant results. In the Motor case, we
see higher performance of the results with the ensemble strategy with a Hyper-
volume (HV) of 1.795 and in the U-Bend case, we observe the best results with
the CNN surrogate and an HV of 1.157. The values of the HV are calculated using
normalized target values and reference points of the maximal values present in
the original databases.

6 Conclusion
Our investigation underscores the efficacy of employing diverse machine learn-
ing surrogate models across heterogeneous engineering use cases. In particular,
 MO-Optimisation Through ML-Supported Multiphysics Simulation 311

our analysis reveals that although the Motor dataset exhibits commendable per-
formance with MAPE values below 1.9%, the U-Bend dataset poses greater
challenges, evident in higher MAPE values. At the hand of a feature importance
analysis and partial dependency plots, we determined key parameters signifi-
cantly influencing output values within both datasets. Validating the solution
candidates against validated values using numerical simulations confirms the effi-
cacy of the surrogate models. Despite higher complexity in the U-Bend case with
some deviations attributed to the scarcity of training data in certain regions, our
approach achieves efficient results. It showcases the versatility of our pipeline in
addressing diverse optimization tasks and the evaluation strategy can be applied
in a general way in future work to multiple use cases. Future investigations will
focus on iteratively refining the optimization cycle and expanding the dataset to
enhance model performance, the robustness of our methodology and its applica-
bility in addressing complex engineering problems.

Acknowledgement. We kindly acknowledge funding by the Federal Ministry for Eco-

nomic Affairs and Climate Action (BMWK) within the project ”KITE: KI-basierte
Topologieoptimierung elektrischer Maschinen” (#19I21034B). We extend our special
thanks to Thomas Menne3 for simulating solution candidates and providing feedback
on the Motor case.

References
1. Bhatnagar, S., Afshar, Y., Pan, S., et al.: Prediction of aerodynamic flow fields
using convolutional neural networks. Comput. Mech. (2019)
2. Chen, T., Guestrin, C.: XGBoost: a scalable tree boosting system. In: Interna-
tional Conference on Knowledge Discovery and Data Mining (SIGKDD). ACM,
San Francisco California USA (2016)
3. Choi, M., Choi, G., Bramerdorfer, G., Marth, E.: Systematic development of a
multi-objective design optimization process based on a surrogate-assisted evolu-
tionary algorithm for electric machine applications. Energies 16, 392 (2023)
4. De Bézenac, E., Pajot, A., Gallinari, P.: Deep learning for physical processes:
incorporating prior scientific knowledge. J. Stat. Mech. Theory Exper. (2019)
5. Deb, K., Pratap, A., Agarwal, S., Meyarivan, T.: A fast and elitist multi objective
genetic algorithm: NSGA-II. IEEE Trans. Evol. Comput. (2002)
6. Decke, J., Gruhl, C., Rauch, L., Sick, B.: DADO – Low-cost query strategies for
deep active design optimization. In: 2023 International Conference on Machine
Learning and Applications (ICMLA), pp. 1611–1618. IEEE (2023)
7. Decke, J., Schmeißing, J., Botache, D., Bieshaar, M., Sick, B., Gruhl, C.: NDNET:
a unified framework for anomaly and novelty detection. In: Schulz, M., Trinitis,
C., Papadopoulou, N., Pionteck, T. (eds.) Architecture of Computing Systems:
35th International Conference, ARCS 2022, Heilbronn, Germany, September 13–15,
2022, Proceedings, pp. 197–210. Springer International Publishing, Cham (2022).
https://doi.org/10.1007/978-3-031-21867-5 13
8. Decke, J., Wünsch, O., Sick, B.: Dataset of a parameterized U-bend flow for deep
learning applications. Data Brief 109477 (2023)
9. Derouiche, K., Garois, S., Champaney, V., et al.: Data-driven modeling for multi-
physics parametrized problems-application to induction hardening process. Metals
(2021)
312 D. Botache et al.

10. El Haber, G., Viquerat, J., Larcher, A., et al.: Deep learning model to assist mul-
tiphysics conjugate problems. Phys. Fluids (2022)
11. Feurer, M., Klein, A., Eggensperger, K., et al.: Efficient and robust auto-
mated machine learning. In: Advances in Neural Information Processing Systems
(NeurIPS), pp. 2962–2970 (2015)
12. Fonseca, C., Paquete, L., Lopez-Ibanez, M.: An improved dimension-sweep algo-
rithm for the hypervolume indicator. In: International Conference on Evolutionary
Computation, IEEE, Vancouver, BC, Canada (2006)
13. Groen, D., Zasada, S.J., Coveney, P.V.: Survey of multiscale and multiphysics
applications and communities. Comput. Sci. Eng. 16, 34–43 (2014)
14. Hachem, E., Ghraieb, H., Viquerat, J., et al.: Deep reinforcement learning for the
control of conjugate heat transfer. J. Comput. Phys., 110317 (2021)
15. Ishibuchi, H., Masuda, H., Tanigaki, Y., Nojima, Y.: Modified distance calculation
in generational distance and inverted generational distance. In: Gaspar-Cunha,
A., Henggeler Antunes, C., Coello, C.C. (eds.) Evolutionary Multi-Criterion Opti-
mization: 8th International Conference, EMO 2015, Guimarães, Portugal, March
29 –April 1, 2015. Proceedings, Part II, pp. 110–125. Springer International Pub-
lishing, Cham (2015). https://doi.org/10.1007/978-3-319-15892-1 8
16. Karniadakis, G.E., Kevrekidis, I.G., Lu, L., et al.: Physics-informed machine learn-
ing. Nature Rev. Phys. 3, 422–440 (2021)
17. Keyes, D.E., McInnes, L.C., Woodward, C., et al.: Multiphysics simulations: chal-
lenges and opportunities. Int. J. High Performance Comput.Appl. pp. 4–83 (2013)
18. Morimoto, M., Fukami, K., Zhang, K., et al.: Convolutional neural networks for
fluid flow analysis: toward effective metamodeling and low dimensionalization.
Theor. Comput. Fluid Dyn. (2021)
19. Oh, S., Jung, Y., Kim, S., et al.: Deep generative design: integration of topology
optimization and generative models. J. Mech. Design (2019)
20. Parekh, V., Flore, D., Schöps, S.: Variational autoencoder-based metamodeling for
multi-objective topology optimization of electrical machines. IEEE Trans. Magnet-
ics, 1–4 (2022)
21. Pedregosa, F., Varoquaux, G., Gramfort, A., et al.: Scikit-learn: machine learning
in Python. J. Mach. Learn. Res., 2825–2830 (2011)
22. Pierrot, T., Richard, G., Beguir, K., Cully, A.: Multi-objective quality diversity
optimization. In: Proceedings of the Genetic and Evolutionary Computation Con-
ference, pp. 139–147 (2022)
23. Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: a
deep learning framework for solving forward and inverse problems involving non-
linear partial differential equations. J. Comput. Phys., 686–707 (2019)
24. Sener, O., Koltun, V.: Multi-task learning as multi-objective optimization. In:
Advances in Neural Information Processing Systems (NeurIPS) (2018)
25. Sharma, P., Chung, W.T., Akoush, B., Ihme, M.: A review of physics-informed
machine learning in fluid mechanics. Energies, 2343 (2023)
26. Zhao, Q., Hastie, T.: Causal interpretations of black-box models. J. Bus. Econ.
Stat. (2021)