<?xml-model href='http://www.tei-c.org/release/xml/tei/custom/schema/relaxng/tei_all.rng' schematypens='http://relaxng.org/ns/structure/1.0'?><TEI xmlns="http://www.tei-c.org/ns/1.0">
	<teiHeader>
		<fileDesc>
			<titleStmt><title level='a'>Accelerating Electric-magnetic machine simulation using the Fourier Neural Operator (FNO)</title></titleStmt>
			<publicationStmt>
				<publisher>American Society of Mechanical Engineers (ASME)</publisher>
				<date>08/17/2025</date>
			</publicationStmt>
			<sourceDesc>
				<bibl> 
					<idno type="par_id">10634656</idno>
					<idno type="doi"></idno>
					
					<author>Zhengke Guo</author><author>Xianfeng David Gu</author><author>Shikui Chen</author>
				</bibl>
			</sourceDesc>
		</fileDesc>
		<profileDesc>
			<abstract><ab><![CDATA[Electrical machines traditionally rely on Finite Element Analysis (FEA) to evaluate or simulate their properties by solving the associated partial differential equations (PDEs). However, FEA is computationally costly, which limits its capability for rapid design iteration and real-time simulations. While recent surrogate models such as Physics-Informed Neural Networks (PINNs) have shown promise, they often suffer from slow convergence and scalability issues in complex geometries. In this paper, we propose the use of the Fourier Neural Operator (FNO) as a resolution-invariant surrogate model to significantly reduce the computation time required for FEA-based PDE solutions in electric machines. Previous research has demonstrated the FNO’s ability to learn mappings for time-sequence problems by approximating operators between function spaces. Building on this, we present a methodology to directly predict the later-state electromagnetic fields of a rotating interior permanent magnet (IPM) motor based on its earlier-stage data by approximating the underlying operator that governs these transitions.Our framework enables full-geometry modeling without relying on segmentation, preserving accuracy while dramatically improving computational efficiency. The model was trained and validated on an FEA dataset with multiple boundary conditions and motor configurations, demonstrating strong generalization across different designs and resolutions. Experimental results show that the proposed FNO method achieves a significant reduction in computational time compared to traditional FEA simulations while maintaining an acceptable level of accuracy. This study highlights the potential of neural operators for accelerating electromagnetic simulations, enabling faster design iterations and offering new possibilities for real-time and optimization-based applications in electric machine design.]]></ab></abstract>
		</profileDesc>
	</teiHeader>
	<text><body xmlns="http://www.tei-c.org/ns/1.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance" xmlns:xlink="http://www.w3.org/1999/xlink">
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="1">Introduction</head><p>Traditionally, physical evaluations in machinery design or other simulation problems are conducted using conventional methods, such as Finite Element Analysis (FEA) <ref type="bibr">[1,</ref><ref type="bibr">2]</ref> and Computational Fluid Dynamics (CFD) for fluid problems <ref type="bibr">[3,</ref><ref type="bibr">4]</ref>.</p><p>In electromagnetic system simulations, these methods typically focus on specific physical properties and require solv-ing complex partial differential equations (PDEs), which can be computationally expensive, especially for large-scale 2D or 3D problems.</p><p>Recent advances in neural network models offer an alternative approach to overcome this computationally expensive process. Researchers have begun to explore neural networks as surrogate solvers for complex physical simulations.</p><p>Early applications of neural networks to physical simulation and inverse problems utilized classic architectures such as Convolutional Neural Networks (CNNs) and Residual Networks (ResNets) <ref type="bibr">[5]</ref>. Li et al. <ref type="bibr">[6]</ref> proposed a recurrent network framework to efficiently reconstruct high-resolution images in nonlinear EM inverse scattering problems that can replace the traditional method. These early works demonstrated the potential of deep learning for surrogate modeling in computational physics.</p><p>More recently, Physics-Informed Neural Networks (PINNs) <ref type="bibr">[7]</ref> have become a popular approach for solving partial differential equations (PDEs) <ref type="bibr">[8]</ref>. By incorporating physical laws into the neural network's loss function, PINNs take PDE constraints and boundary conditions into account during training. Samaniego et al. <ref type="bibr">[9]</ref> proposed a PINN framework that minimizes the total energy of the system, enabling the solution of energy-based PDEs in mechanics problems. Lu et al. <ref type="bibr">[10]</ref> introduced methods for imposing Dirichlet and periodic boundary conditions within the PINN framework, extending its applicability to new types of PDEs. Yuan et al. <ref type="bibr">[11]</ref> demonstrated the use of PINNs to solve integro-differential equations (IDEs), highlighting their flexibility. Son et al. <ref type="bibr">[12]</ref> developed a PINN-based surrogate model for evaluating the electromagnetic responses of electric motors.</p><p>Although PINNs offer significant advantages, they also face challenges in solving complex geometries and high-dimensional problems. The physical governing equations act as soft constraints during training, and adding more terms to the loss function often increases optimization difficulty. PINNs frequently suffer from slow convergence and sensitivity to hyperparameter selection, which can limit their scalability and robustness. Wang et al. <ref type="bibr">[13]</ref> addressed these challenges by reintroducing elements and nodes into the PINN framework to improve accuracy and convergence, simulating certain aspects of the finite element method (FEM). However, this approach compromises the meshfree advantage typically associated with PINNs.</p><p>However, in solving complex structure problems, PINN is computationally difficult and is hard to converge. And the physical governing equations work more as a soft constraint. Adding more terms to loss functions will sometimes further increase the difficulty of training.</p><p>An alternative approach is provided by Deep Operator Networks (DeepONets) <ref type="bibr">[14]</ref>, which learn mappings between input and output functions and have shown promise in solving PDEs and ODEs. A DeepONet consists of two networks: the Branch Network, which encodes the input function sampled at fixed sen-sor locations, and the Trunk Network, which encodes query locations. The inner product of their outputs predicts the solution at arbitrary points. DeepONets have demonstrated success in solving nonlinear dynamical systems <ref type="bibr">[14]</ref>. Yin et al. <ref type="bibr">[15]</ref> proposed a multiscale framework integrating DeepONet with FEM, while Goswami et al. <ref type="bibr">[16]</ref> incorporated a minimal energy criterion into DeepONet to solve brittle fracture PDEs. Goswami et al. <ref type="bibr">[17]</ref> introduced Physics-Informed DeepONets. Lu et al. <ref type="bibr">[18]</ref> introduced a multi-fidelity DeepONet framework to solve nanoscale heat transport problems and further combined the method with genetic algorithms (GA) and topology optimization.</p><p>Koenig et al. and Shukla et al. <ref type="bibr">[19,</ref><ref type="bibr">20]</ref> use the newly introduced Kolmogorov-Arnold Network <ref type="bibr">[21]</ref> to solve ODEs and PDEs.</p><p>Neural operators <ref type="bibr">[22]</ref> have emerged as a powerful class of machine learning methods. They approximate operators that learn mappings from function inputs, such as initial conditions or boundary conditions, to function outputs, such as solutions to partial differential equations (PDEs), in a resolution-invariant manner. Among these, the Fourier Neural Operator (FNO) <ref type="bibr">[23]</ref> is notable for its use of fast Fourier transforms in each layer, enabling globally non-local representations and efficient training for PDE-related tasks. Kovachki et al. <ref type="bibr">[24]</ref> proved that the FNO is a universal approximator capable of learning any continuous operator and discussed its potential for learning operators associated with PDEs.</p><p>The follow-up research <ref type="bibr">[25]</ref> introduced a deformation-based approach to handle arbitrary geometries instead of using direct interpolation. The author introduces a few neural network layers to transform mesh-grid data into the neural network. Then, some inverse transformation layer is used to return the data to the input mesh frame. Cao et al. <ref type="bibr">[26]</ref> proposed a neural operator based on the Laplace transform, replacing the previous Fourier transform. <ref type="bibr">Guibas et al. [27]</ref> proposed a combination of transformer models with a Fourier Neural Operator as a token mixer named Adaptive Fourier Neural Operators. Since transformers <ref type="bibr">[28]</ref> are powerful in big-data modeling, Pathak et al. <ref type="bibr">[29,</ref><ref type="bibr">30]</ref> used Adaptive Fourier Neural Operators to accelerate global weather forecasts, where variables such as surface wind speed, precipitation, and atmospheric water vapor can be predicted using neural networks. Compared to the PINN method, where the physical formulation data work as a loss function. The physical information is hidden in the dataset when using the neural operator. Hao et al. <ref type="bibr">[31]</ref> proposed a large-scale neural operator transformer framework with the potential to solve physical field PDE. Goswami et al. <ref type="bibr">[17]</ref> conducted a comprehensive review of both Neural Operators and DeepONets, and proposed a Physics-Informed Neural Operator framework inspired by both of them.</p><p>Currently, to reduce computational costs, ordinary simulations leverage the motor's symmetrical structure by focusing on only one segment or sector instead of the entire model. While this approach indeed lowers computational requirements, it poses challenges for analyzing thermal behavior and mechanical stress.</p><p>By applying a neural network to accelerate FEA simulations, it becomes possible to simulate the entire motor geometry while still benefiting from cost savings. This approach maintains the thoroughness of a full-sector model and addresses the limitations that arise when the motor is partitioned into segments.</p><p>This paper focuses on applying the baseline Fourier Neural Operator to accelerate FEA simulations of interior permanent magnet (IPM) motors. The Fourier Neural Operator (FNO) was selected for this task due to its demonstrated effectiveness in modeling continuous complex physical systems and its inherent ability to generalize across different spatial resolutions. These characteristics are valuable for electric motor simulations involving diverse geometric and electromagnetic configurations. Unlike recent work emphasizing large-scale deep learning architectures such as transformers, our approach leverages relatively small models based on FNOs for efficiency and scalability. This study is organized as follows: Section 2 reviews the theoretical background of neural operators; Section 3 describes the training process and implementation of FNOs for electric motor simulations; Section 4 presents the results and analysis; and Section 5 concludes the paper with discussions on the method's potential and planned future work.</p><p>The main contributions of this paper are summarized as follows:</p><p>1. We propose a novel framework that applies Fourier Neural Operators (FNOs) to accelerate electromagnetic field simulation in electric motors, enabling fast simulations across rotor positions while maintaining high accuracy. 2. We introduce a preprocessing pipeline that transforms data from the circular motor domain to a uniform rectangular grid, allowing compatibility with FNO input formats and preserving geometric structure.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2">Neural Operator Theory and Fourier Spectral Implementation</head><p>A Neural Operator is a neural network architecture designed to learn mappings between infinite-dimensional function spaces from finite datasets. Unlike traditional neural networks, such as convolutional neural networks (CNNs), which operate on fixedsize, finite-dimensional data, neural operators are trained on finite data pairs but aim to learn the underlying operator that maps between entire function spaces. This capability demonstrates their potential in solving complex partial differential equation (PDE) systems.</p><p>One of the key features of neural operators is their discretization invariance. Once trained, they can generalize across different resolutions within the same function space, allowing them to process inputs and outputs with varying discretizations without the need for retraining. This makes neural operators highly flexible and efficient for applications involving complex physical systems.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.1">Operator Learning Problem Formulation</head><p>The problem setting of an operator is set in the following way. Let D &#8834; R d be a bounded, open domain. The input-output pair finite dataset can be treated as input function A = A(D; R d a ) and output U = U(D; R d u ). Where both A and U is function space defined on D in separable branch space R d a ,R d u . Here exists a true operator G &#8224; A &#8594; U. We are looking for operators to approximate it by constructing a mapping:</p><p>The target is to find a parametric operator G &#952; , with parameters &#952; located in finite-dimensional parameter space</p><p>The parametric operator G &#952; is constructed using the neural operator in the research. For a neural network to learn, we will need to obtain the datasets from function space A and U. we have observation a j , u j N j=1 . where a j &#8764; &#181; is an independent and identically distributed (i.i.d.) sequence sampled from distribution &#181; over the function space A. The paired outputs are given by u j = G &#8224; (a j ).</p><p>Since the a j and u j here are an observation of the function space A and U. Instead of approximating a relationship between two sets of discrete data points, as in traditional neural network learning, the neural operator learns a mapping between functions. Therefore, this relationship is discretization-invariant. However, considering that we are trying to approximate the true operator G &#8224; using G &#952; , this approximation may fail in practice when there is a significant difference between the resolutions of the training data and the application data.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.2">General Neural Operator Architecture</head><p>Considering the above problem setting <ref type="bibr">[22]</ref>, the neural operator is constructed in the form of an iterative architecture, which can be written as:</p><p>Here, P : R d a &#8594; R d v is a local lifting operator that maps the input a(x) into a branch space R d v , and Q : R d v &#8594; R d u is a local projection operator that maps from branch space R d v to the output branch space R d u where u(x) located.</p><p>There are L 1 , . . . , L T non-linear operator layers. Each layer L &#8467; : v &#8594; L &#8467; (v) acts on a function v in the function space (D; R d v ) and produces another function v in the same space. Each layer L &#8467; (v) can be expressed as:</p><p>where W &#8467; is the pointwise local linear weight, b &#8467; (x) is the pointwise local bias, and &#963; is a non-linear activation function. The most important component is the non-local linear operator K.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.3">Fourier Neural Operator</head><p>The Fourier Neural Operator (FNO) implements the operator K through spectral methods. While traditional neural network architectures process information locally in the spatial domain, FNOs operate in the frequency domain, enabling them to capture global dependencies more efficiently. By projecting the input function into the Fourier space, applying learned transformations to the frequency components, and then projecting back to the spatial domain, FNOs can approximate complex non-local operations with fewer parameters than conventional architectures. This spectral approach offers two key advantages: (1) capturing long-range dependencies through global operations in the frequency domain, and (2) achieving discretization invariance, as the underlying continuous functions are approximated rather than their discrete samples.</p><p>As introduced in the previous section, the operator K can be defined as an integral operator <ref type="bibr">[22,</ref><ref type="bibr">24]</ref>. In this research, we specifically utilize the Fourier Neural Operator (FNO) <ref type="bibr">[23]</ref>, where K is implemented using Fourier layers. This formulation brings the neural operator into the Fourier domain. For each layer L &#8467; , the layer can be formulated as:</p><p>where F and F -1 are the Fourier transform and inverse Fourier transform, respectively. P &#8467; (k) are learnable coefficients in the Fourier space, parameterized by the frequency mode k.</p><p>The mode k plays a critical role in the FNO network, as it determines the truncation of the Fourier series. In engineering applications, a hyperparameter k max is typically set, and only modes with k &#8804; k max are retained. This means that only the lower frequency components in the Fourier domain are preserved, which reduces model complexity while maintaining essential information.</p><p>The complete structure of a Fourier Neural Operator model is illustrated in Figure <ref type="figure">1</ref>.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="2.4">FNO-Based Electromagnetic Field Prediction</head><p>Framework Let us recall the problem setting for the operator. In the proposed method, we aim to use the Fourier Neural Operator (FNO) to find an approximate parametric operator G &#952; that maps the magnetic properties at early rotor angles to their future states at later rotation angles. Specifically, the input function space A represents a high-dimensional function of the electromagnetic properties at earlier rotor angles, while the output function space U represents a high-dimensional function of the electromagnetic properties at later stages. Here, G &#952; is the FNO with parameters &#952; , located in the finite-dimensional parameter space &#920;.</p><p>To approximate the true operator, we acquire functions a(x) from the function space A. The variable x represents the boundary conditions and material properties, which are implicitly encoded within the data points. The function a(x) denotes the electromagnetic field at earlier rotor angles, ranging from &#945; to &#945; + 9 &#8226; . The corresponding output function u(x) represents the predicted electromagnetic field over the subsequent rotor angles, from &#945; + 10 &#8226; to &#945; + 40 &#8226; . The angle &#945; is an arbitrary rotation angle between the rotor and stator of the motor and can take any value between 0 &#8226; and 360 &#8226; .</p><p>This formulation enables the FNO to learn the underlying operator G &#952; , which captures the evolution of the electromagnetic field as the rotor advances through different angular positions.</p><p>In this research, the magnetic vector potential A was selected as the primary physical quantity of dataset observation from function space. Since the motor simulation was conducted in a two-dimensional plane, the only non-zero component of A is along the z-axis. Therefore, A z was used as the dataset for the FNO.</p><p>The selection of A z is motivated by its direct relationship with the magnetic flux density B through the following partial differential equation:</p><p>Considering A z is the only non-zero component of A, the relationship can be further written as:</p><p>Where</p><p>x and B z = 0. This relationship makes A z a convenient and informative parameter for describing the electromagnetic behavior within the cross-sectional domain of the motor. Therefore, the functions a(x) and u(x) consist of sampled A z .</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3">Implementation Methodology for Motor Field Prediction</head><p>When simulating a revolving motor, the rotor typically rotates while the stator remains stationary. In many cases, evaluating the rotor's performance requires a full 360 &#8226; simulation. However, by drawing an analogy from how Fourier Neural Operators (FNOs) handle time-sequence fluid problems, we can treat the electric motor simulation as a sequence of states characterized by different rotor angles rather than as a direct mapping from an initial condition to subsequent states.</p><p>To feed the simulation data into the FNO, the data defined on the circular motor domain is first mapped onto a square matrix. After obtaining the corresponding Cartesian coordinates, the magnetic vector potential values are extracted and formatted appropriately for input to the FNO model.</p><p>The simulation data is resampled into low-resolution datasets. This facilitates training and testing of the FNO's discretization-invariance property, where the network can generalize across different spatial resolutions.</p><p>After training, the FNO model uses data from earlier rotor angles to predict the motor's electromagnetic field distribution at later angles. Consequently, instead of simulating the entire motor from 0 &#8226; to 360 &#8226; , only a limited number of angles need to be computed, significantly reducing the overall computational effort.</p><p>Finally, the predicted magnetic vector potential component A z is used to calculate the corresponding magnetic flux density components B x , B y , and the magnitude &#8741;B&#8741;, completing the evaluation of the motor's electromagnetic behavior.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.1">Simulation Setup and Dataset Creation</head><p>The device used for FEA simulations and neural network training was equipped with an 11th Gen Intel&#174; Core&#8482; i7-11700 CPU @ 2.50GHz, 64GB of RAM, and an NVIDIA GeForce RTX 3070 GPU.</p><p>To train and test the proposed methodology, a dataset was generated using Finite Element Analysis (FEA) of an interior permanent magnet (IPM) motor in COMSOL Multiphysics. The simulated motor consists of a stator and a rotor, with permanent magnets embedded within the rotor at specific angles relative to the radial direction. The magnet placement angle plays a critical role in dataset creation, as it directly influences the electromagnetic behavior of the motor.</p><p>As shown in Figure <ref type="figure">2</ref>, the motor consists of a rotor and a stator. The core structures of both the rotor and stator are made of soft iron, while the coils are composed of copper. The interior permanent magnets are embedded within the rotor. Due to the complexity of the air gap between the rotor and stator, a finer mesh density was applied in this region. In addition, multiple mesh boundary layers were implemented to accurately capture potential strong gradient changes in the electromagnetic fields within the air gap.</p><p>To incorporate multiple boundary conditions into the dataset, three key parameters were selected based on varying key designs and operational parameters of the interior permanent magnet (IPM) motor. The following parameters were considered during dataset generation:</p><p>Permanent Magnet (PM) Angle: The orientation angle of the permanent magnets was varied between 20 &#8226; and 50 &#8226; with one degree as increments. Current Peak Values: The peak amplitude of the input current was set to discrete values of 200 A, 300 A, 400 A, and 500 A. Stator Outer Edge Shape: Three different geometric profiles for the outer edge of the stator were considered: Square, Squircle and Half-circle Rotor-Stator Relative Angle: For each configuration, the relative angle between the rotor and stator was varied continuously from 0 &#8226; to 360 &#8226; with one degree as increments.</p><p>By combining these parameters, a total of 372 motor design combinations were generated, each undergoing a full simulation scenario from 0 degrees to 360 degrees. This comprehensive dataset was created to ensure that the neural network model could generalize across various boundary conditions and motor designs.</p><p>The total simulation time for generating the dataset was 67 hours, 54 minutes, and 44 seconds, with each design taking approximately 10 minutes and 57 seconds to complete.</p><p>After the simulation, a mapping process was implemented</p><p>Copyright &#169; 2025 by ASME  to adapt the simulation data to the input format required by the neural network, which typically expects data on a uniform rectangular grid. Specifically, the data was mapped from a 200 &#215; 200 square grid to a circular domain representing the cross-sectional area of the motor, which includes both the rotor and stator, with an outer radius of 17.5 cm. As a result of this mapping, the air region outside the stator was excluded from the dataset. This transformation ensures a more uniform distribution of grid points across the circular domain and improves the consistency of the data representation <ref type="bibr">[32]</ref>.</p><p>The transformation from the square domain (x, y) to the circular motor domain (u, v) is defined as follows:</p><p>x 2</p><p>x 2 + y 2 when x 2 &#8805; y 2 sgn(y) xy</p><p>Here, (x, y) are the Cartesian coordinates in the square domain, and (u, v) are the corresponding Cartesian coordinates in the circular motor domain. An illustration of the mapping is shown in Figure <ref type="figure">3</ref>.</p><p>After obtaining the Cartesian coordinates of the mapped grid points in the circular domain, the local values of the magnetic properties of interest were extracted at each point. Since the deep learning grid does not coincide with the finite element mesh used in the FEA simulations, interpolation was required during this process to obtain accurate property values at the designated grid points.</p><p>Therefore, the final dataset contains 372 different motor designs, each consisting of 361 simulation steps at a spatial resolution of 200 &#215; 200 grid points and the A z value at the grids. After generating the raw simulation data, the samples were organized into input-output pairs for supervised learning. As mentioned earlier, data from 10 continuous rotor angles were used as the input to predict the motor state over the subsequent 40 degrees of rotation.</p><p>Since 0 &#8226; follows 359 &#8226; , the redundant 360 Due to hardware limitations, only 31 designs were selected from the 372 available motor configurations to construct the final training and testing datasets. The spatial resolution of each data layer was downscaled to 100 &#215; 100. The selected designs were randomly chosen from different combinations of peak current (I pk ) values and stator outer-edge shapes, ensuring that each permanent magnet (PM) angle from 20 &#8226; to 44 &#8226; was represented in the training data. PM angles from 45 &#8226; to 50 &#8226; were reserved exclusively for testing and served as the test dataset to evaluate the model's generalization capability. For the testing data, the original high-resolution 200 &#215; 200 data was preserved to validate the discretization-invariance property of the Fourier Neural Operator (FNO) model. To clarify, for the low-resolution data pairs, the final input fed into the FNO model has a resolution of 10 &#215; 100 &#215; 100, with the corresponding output resolution of 40 &#215; 100 &#215; 100. For the high-resolution data pairs, the input has a resolution of 10 &#215; 200 &#215; 200, and the output has a resolution of 40 &#215; 200 &#215; 200.</p><p>In total, the final dataset consists of 9, 000 low-resolution training data pairs, 2, 160 low-resolution testing pairs, and an additional 2, 160 high-resolution testing pairs.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="3.2">FNO Model Configuration and Training</head><p>The Neural Operator library <ref type="bibr">[33]</ref> was used to build the Fourier Neural Operator (FNO) architecture in this study. For model training, the learning rate was set to 8 &#215; 10 -3 . k max was set to 20. The relative H1 Sobolev norm loss was employed as the training loss function, while both relative L2 and relative H1 losses were used for validating the testing data.</p><p>After testing, to avoid overfitting and maintain good performance in the testing dataset. The total training epoch was set to 30, with a batch size set to 10.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="4">Performance Evaluation and Comparative Analysis</head><p>After training, we evaluated the performance of the model. The training loss is shown in Figure <ref type="figure">4</ref>, where the loss stabilizes around epoch 30, indicating convergence. We observe that the L 2 loss exhibits a noticeable spike at epoch 5 during the early stage of training. As training progresses, the loss gradually stabilizes after epoch 18, reflecting the model's increasing ability to capture the underlying data structure. This early instability might be due to the random initialization of the network parameters, which can lead to large gradient updates in the initial epochs. The average training time per epoch was 125.63 seconds, resulting in a total training time of 3769 seconds. While the training phase incurs an initial computational cost, it is a one-time process. Once trained, the model can predict the electromagnetic field distribution at future rotor angles in real time.</p><p>The only additional major computational time required is for calculating the input data at the initial 10 &#8226; rotor angles using FEA. If the predicted outputs are not reused as new inputs, it is necessary to calculate 80 additional angles using FEA. Specifically, the data from the first 10 &#8226; are used to predict angles from 10 &#8226; to 49 &#8226; . After that, angles from 50 &#8226; to 59 &#8226; are computed and used to predict the next batch from 60 &#8226; to 99 &#8226; . This iterative strategy continues until the full rotation is completed, where predictions are made for angles from 350 &#8226; to 359 &#8226; without requiring further FEA calculations. The total computation time for this approach is approximately 1.5 minutes.</p><p>Compared to the traditional FEA process described in the previous section, where each motor design simulation required approximately 11 minutes, the trained FNO model offers a substantial reduction in computation time. Inference with the FNO enables rapid evaluations across multiple rotor positions, making it particularly suited for design iterations, optimization studies, and potential real-time applications.</p><p>Figure <ref type="figure">5</ref> presents the ground truth, prediction, and absolute error for the first and last output channels of the lowresolution (100 &#215; 100) test data. In Figure <ref type="figure">6</ref>, the corresponding high-resolution (200 &#215; 200) prediction is visualized after being mapped back to the circular motor domain. Upon closer examination, the absolute error in the first output channel is generally smaller than that of the last output channel (channel 39). This is expected, as the last channel corresponds to the furthest time step from the input data. Despite this, the absolute error remains relatively small, demonstrating the model's ability to make accurate predictions even for extended future states.</p><p>With the newly acquired high-resolution predictions of A z , we calculated the corresponding magnetic flux density B in the interior permanent magnet (IPM) motor, as shown in Figure <ref type="figure">7</ref>. To perform this computation, the high-resolution A z data was first interpolated onto a uniform mesh grid within the motor's cross-sectional domain. The corresponding B x and B y components were then derived from A z . These predicted B x and B y values were subsequently compared to the reference results obtained from the finite element analysis (FEA) simulation.</p><p>The error in the predicted B-field is larger than that observed in the A z predictions. The primary errors occur at the geometric boundaries between different materials. This error is more obvious in the later output channel 39, which represents the FNO's furthest prediction step. Some of these errors may be attributed to interpolation inaccuracies that arise during the transformation between the model's grid and the evaluation mesh.</p><p>A significant portion of the error is concentrated along the material boundaries. This may be due to the inherent smoothness of the FNO output, which makes it difficult to capture sharp gradient transitions. In particular, the Fourier layers in the FNO architecture deliberately filter out high-frequency components, resulting in the loss of fine detail. While this loss of highfrequency modes has minimal impact on the A z predictions, it becomes problematic during the curl computation of A to derive B. The smoothing effect leads to amplified errors along material interfaces where sharp changes in field values occur.</p><p>Despite these challenges, and considering our model was trained on low-resolution data.The overall distribution and trends of the predicted B-field align well with the FEA results. The operator has successfully approximated the mapping between an early stage of the Motor electromagnetic field to the later stage, and the Discretization of the proposed Fourier neural operator is shown. The proposed pipeline demonstrates significant potential for accelerating motor FEA simulations and providing an efficient alternative to traditional PDE solvers.</p></div>
<div xmlns="http://www.tei-c.org/ns/1.0"><head n="5">Conclusion and Discussion</head><p>This paper proposes a novel methodology for accelerating electromagnetic finite element analysis simulations and predicting the electromagnetic properties of an interior permanent mag-    net motor. We discuss the concept of neural operators and use the Fourier neural operator to approximate the mapping. Through careful dataset generation and sampling strategies, we trained and validated the model, with successful predictions of the magnetic vector potential A z and derived magnetic flux density B. Demonstrating its capability to perform high-fidelity predictions while significantly reducing simulation time compared to traditional FEA approaches.</p><p>Despite the success of the simulation. The current study relies on interpolation in the complicated Mesh space. Considering the complicated geometry of the motor. This raised some concern when simulating the air gap between the rotor and the stator.</p><p>The current interpolation step introduces potential sources of error. In particular, the sharp gradients and discontinuities in the air gap region make accurate field reconstruction more challenging. As mentioned earlier, Li et al. <ref type="bibr">[25]</ref> introduced a deformationbased approach to handle arbitrary geometries instead of using direct interpolation to obtain a uniform dataset structure. Using a transformation process to transform the complex mesh structures into the neural network could further improve the framework's robustness. Additionally, the current neural operator still requires the use of the ordinary FEA method. Finding neural operator models capable of directly mapping from a function of input boundary conditions to a function of output solutions might still be an option, potentially eliminating the reliance on conventional FEA preprocessing.</p><p>The current methodology has been successfully applied to predicting A z and the derived magnetic flux density B. These results demonstrate its potential for predicting additional electromagnetic properties and might eventually apply to predicting motor performance metrics, such as output torque. Furthermore, this approach shows promise in the field of topology optimization, where forward simulations are often computationally expensive. Since topology optimization requires iterative simulations as material distributions evolve, accelerating these simulations with neural operators could significantly reduce computation time and enable more efficient optimization processes <ref type="bibr">[34,</ref><ref type="bibr">35]</ref>.</p><p>In summary, the proposed framework demonstrates both the effectiveness and flexibility in accelerating electromagnetic simulations. Future work will aim to address current limitations, improve geometric handling, and extend the applicability of this approach to more complex electromagnetic problems and multiphysics simulations.</p></div><note xmlns="http://www.tei-c.org/ns/1.0" place="foot" xml:id="foot_0"><p>Copyright &#169; 2025 by ASME</p></note>
		</body>
		</text>
</TEI>
