Conference Agenda

The accurate prediction of magnetic susceptibility in rare earth (RE) compounds presents a significant challenge due to their intricate electronic structures and strong electromagnetic interactions. This study leverages machine learning (ML) techniques as powerful tools to model and predict the magnetic susceptibilities of RE compounds, focusing on their electromagnetic properties. By employing ML algorithms such as linear regression, decision trees, random forests, XGBoost, and neural networks, we aim to address the computational challenges associated with traditional numerical methods for solving electromagnetic field problems in materials with partially filled 4f orbitals. Analysing a dataset of RE halides, oxides, sulfates, and sulfides, we demonstrate the robustness of ensemble methods like random forests and XGBoost in delivering accurate predictions even with limited data. The research highlights the potential of integrating ML into the computational analysis of magnetic material properties, offering a streamlined and data-driven alternative to traditional approaches. By improving the efficiency and accuracy of susceptibility predictions, this study contributes to advancements in computational electromagnetics, with implications for the design of materials tailored for specific magnetic applications.

This paper proposes an optimal design method for surface-mounted permanent magnet synchronous motors (SPMSM) based on electromagnetic performance prediction using a convolutional neural network (CNN)-based regression model enhanced with attention mechanisms. By capturing spatial structures and local patterns, CNNs effectively extract key features from motor designs, making them well-suited for performance prediction. The integration of attention mechanisms further improves the model by assigning greater weights to critical design features, enhancing prediction accuracy and learning efficiency. Additionally, attention mechanisms enable the identification of relationships among key elements, addressing the limitations of localized learning in conventional CNN models. In this study, the effectiveness of attention mechanisms is evaluated, and a CNN-based regression model incorporating these mechanisms is developed. The proposed regression model is then applied to optimize the design of an SPMSM.

Electric vehicle (EV) propulsion motors operate under different conditions, causing input currents to vary based on MTPA and MTPV control criteria. These changes in input currents result in non-linear variations in air-gap electromagnetic force harmonics (AEFHs). To reduce vibrations across various operating points, non-linear characteristics must be considered during the design stage. However, designing to reduce AEFHs early in the design stage across various operating points incurs high computational costs, as it requires d-q flux linkage information for various d-q current combinations to determine input conditions and calculate AEFHs based on control logic. To address this, we employ a neural network (NN) model to predict d-q flux linkages and AEFHs based on geometric parameters and d-q currents. The trained NN uses currents derived from MTPA and MTPV control logic along with voltage equations based on the predicted d-q flux linkages. This enables the prediction of motor performance and AEFHs across various operating conditions, facilitating computationally efficient motor design.

Axial flux permanent magnet synchronous motors (AFPMs) have high torque density due to its low magnetic reluctance for stator core, so that AFPMs are used for future mobility. For the traction system in future mobility, high efficiency of AFPMs is required for its long all-electric range. In order to design high efficiency AFPMs, accurate calculation of pulse-width modulation (PWM) induced iron loss is necessary, but it requires high computation cost because of its small time step and complex 3-D magnetic flux path of AFPMs. Therefore, it is hard to construct a surrogate model of PWM-induced iron loss according to the design variables of AFPMs. In this paper, we propose a deep transfer learning-based surrogate modeling process of PWM-induced iron loss for AFPMs with dramatically reduced computation cost. The proposed method incorporates a small amount of PWM-induced iron loss of AFPMs calculated by using 3-D finite element analysis (FEA) and a large amount of sinusoidal current-based iron loss calculated by using 2-D FEA. By using the proposed method, PWM-induced iron loss of AFPMs can be predicted with much reduced computation cost, so that design of high efficiency AFPMs considering PWM harmonics can be allowed. The effectiveness of proposed method will be validated through simulation and experimental results.

This paper proposes a transfer learning approach based on an analytic method to efficiently estimate the electromagnetic characteristics of a Surface-mounted Permanent Magnet Synchronous Motors (SPMSM). The characteristics of SPMSM are determined by geometric variables and input current conditions and the methods for calculating these characteristics include analytic methods and Finite Element Analysis (FEA). Analytic methods offer rapid computation and are suitable for large scale data but struggle with nonlinear characteristics, while FEA provides high accuracy but requires longer computation time. To overcome these limitations, this paper proposes a method that utilizes an analytic approach to rapidly generate large scale data, trains a deep learning model based on this data, and then applies transfer learning with a small amount of FEA data to enhance the accuracy of the performance prediction model while reducing computation time. The proposed method has been confirmed to rapidly and accurately estimate the electromagnetic characteristics of SPMSM according to geometric variables and current, and by complementarily utilizing the advantages and disadvantages of both analytic methods and FEA, it significantly reduces computation time during the design process.

This paper presents a novel method for estimating impact locations of radar-homing anti-ship missiles based on the radar cross-section characteristics of naval ships. The study employs two different approaches: one using multivariate distribution and the other data-driven machine learning techniques. The multivariate distribution method analyzes the correlation between variables, such as azimuth angle, elevation angle, and distance, enabling quick trend identification. In contrast, the machine learning approach provides more precise impact point predictions but requires significantly longer data analysis and model training time. Results demonstrate that the predicted impact locations using machine learning closely match the equivalent scattering center, whereas the multivariate distribution approach reveals a broader spread of impact points. Future work will include detailed mathematical formulations for multivariate distribution and covariance, as well as the complete machine learning algorithm.

Optimizing electromagnetic performance and mechanical stress is essential in designing high-speed traction motors for electric vehicles. Although research on the motor design using deep learning (DL)-based optimization is actively conducted, it often requires significant data sampling time to achieve accurate predictions. This study proposes a hybrid DL–finite element analysis (FEA) approach to address these challenges and accelerate multi-physics optimization. FEA-generated data from each optimization iteration is used to train the DL model, whose accuracy is evaluated via root mean square error (RMSE) and R-squared metrics. Once the model meets specified accuracy thresholds, the predicted mechanical stress is compared with constraints to narrow the motor design space, eventually replacing mechanical stress FEA at higher accuracy levels. Consequently, the proposed method significantly reduces the time required for multi-physics optimal design.

Recently, there are a lot efforts in developing machine learning and deep learning methods for the prediction of electric motor performances. In particular, torque profile for a given motor design, including cogging torque and torque ripple, are challenging for surrogate models to achieve high prediction accuracy. One promising approach is to represent a motor design as a 2d image, and utilize deep learning models that found success in image recognition and classification tasks for motor performance prediction. A number of deep convolutional neural networks (CNNs) have been adopted for motor applications previously, while more recently Vision Transformer (ViT) models are gaining interests. In this paper, we evaluate multiple deep CNN and ViT models on two datasets of interior permanent magnet motors, and show that ViT based models can achieve superior accuracy for cogging torque prediction compared with CNN based models, and can be jointly trained on the two different datasets and still provides prediction with low error.

Electric motor design and optimization is a complex process that must simultaneously satisfy various performance indicators such as output, efficiency, and minimization of cogging torque. Finite Element Analysis (FEA) is widely used for this purpose, but its high computational cost and time requirements pose limitations for large-scale design problems. In this study, Kriging and Deep Neural Network (DNN) are compared as surrogate models for predicting the performance of Permanent Magnet Synchronous Motors (PMSMs). Design variables were extracted using Latin Hypercube Sampling (LHS), and FEA was utilized to derive electromagnetic characteristics such as torque, torque ripple, d- and q-axis inductances. The Kriging and deep learning models were evaluated based on data utilization efficiency and prediction accuracy, and the strengths and weaknesses of each method for motor design problems are analyzed. This study proposes practical approaches for effectively utilizing surrogate models to enhance the efficiency of motor design process.

Surface-mounted permanent magnet synchronous motors (SPMSMs) for robot actuators are typically designed in a pancake form due to the limited space within actuator systems. In order to control the robot manipulator in precision, vibration characteristic is one of the main issue in designing the SPMSMs for robot actuators. As the radial and tangential electromagnetic force which are considered as an excitation force in stator are affected by the air-gap magnetic flux density, electromagnetic analysis considering the axial leakage flux is important in pancake formed SPMSMs. Thus, 3-D finite element analysis (FEA) is necessary to calculate electromagnetic excitation force density (EEF) for SPMSMs, but it requires high computation cost. Therefore, analytical model-based deep transfer learning is proposed in this paper. The proposed method incorporates a large amount of analytical model-based radial and tangential electromagnetic force and a small amount of 3-D FEA-based electromagnetic force, so that the EEFD can be redicted accurately with reduced computation cost. By using the proposed method, the low vibration and high performance SPMSMs for robot actuator can be designed rapidly considering the axial leakage flux. The results of analysis using the proposed method will be validated through simulation and experimental results.

Electrically excited synchronous motors (EESMs) have high degree-of-freedom in controlling the current vector for armature winding, field current, and rotational speed, so that EESMs can be operated in wide torque and speed range. In general, 2-D finite element analysis (FEA) is used for analysis of efficiency map for EESMs because of its high accuracy. Therefore, a lot of 2-D FEA should be performed repeatedly according to the current vector, field current, and rotational speed to calculate the flux linkage and iron loss of the EESMs. This leads to increased costs in designing high-efficiency EESMs, so the development of an analysis technique for efficient efficiency map is necessary. This paper proposes an equivalent magnetic circuit (EMC) based deep transfer learning for efficient analysis of efficiency map for EESMs. EMC allows us to calculate flux-linkage and iron loss of EESMs rapidly, but correction method should be accompanied because of its relatively low accuracy compared to 2-D FEA. The proposed method incorporates a large amount of EMC-based flux linkage and iron loss, and a small amount of 2-D FEA-based flux-linkage and iron loss so that the efficiency map for EESM can be predicted accurately with reduced computation cost. The proposed method is applied to analysis of the efficiency map, and its validity is verified through simulation and experimental results.