Conference Agenda

Physics-informed neural network (PINN) is a method for training neural networks by constructing a loss function based on the error in the initial-boundary value problems for partial differential equations. This study focuses on the application of PINN to solve the axisymmetric magnetostatic field problems. PINN is learned using a point set but does not use connection information between points like the Yee’s grid or edge elements. Therefore, interface conditions for electromagnetic fields are not expected to be satisfied naturally. To solve this issue, this study proposes a PINN algorithm based on domain decomposition methods that divides the analysis domain into conductor and non-conductor and satisfies the continuity of the magnetic flux density and magnetic field strength at the boundaries between subdomains by an iterative method.

This paper presents an efficient parameter-topology hybrid optimization method based on a data-driven approach. In the present method, the shapes of the magnets are represented by geometric parameters, and those of the iron cores are determined through topology optimization. Because the magnet and topology parameters are related to each other, their simultaneous optimizations lead to difficult optimization category. The present method relaxes this issue by making the search interval of the magnet parameters small. To find the effective search space for magnet parameters, reinforcement learning is used, that is, effective shapes of the magnets are estimated through the data generated by computationally light evaluations. The reinforcement learning utilizes the data and estimates the effective magnet parameters. Moreover, thanks to the use of the data-driven approach, additional optimization requirements can be processed more quickly.

In this paper, active learning (AL) is introduced to accelerate the surrogate modeling for electromagnetic device, requiring fewer samples to achieve higher accuracy. A 3D inductor model with nonlinear magnetic material, described by 8 design parameters is used for the prediction of the inductances under 5 different current excitations, where random forest regressor (RFR) is employed as the surrogate model, trained via AL and random sampling (RS). The results indicate that the models trained by AL-selected samples outperform those trained by RS when data sets are limited. Meanwhile, AL demonstrates smoother learning curves, indicating enhanced robustness.

The finite element method often uses quadrilateral meshes, which are better than triangular meshes for calculating results. However, it is still hard to automatically generate quadrilateral meshes. This study proposes a system that automatically generates meshes using large language models. We show that state-of-the-art large language models can successfully generate high-quality quadrilateral meshes for simple two-dimensional shapes. This suggests that there is potential for automated structural design workflows.

The novel efficient parameter optimization algorithm for interior permanent magnetic motors that prevents shape collapse is proposed. When the shape collapse is detected, the proposed method generates an individual by perturbing a non-collapsed individual. This process reduces the rate of shape collapse compared to randomly generating new individuals. It is numerically shown that the proposed method works 21 times faster than the conventional method while maintaining the optimization performance.

Performances of two different surrogate model assisted optimizations are compared. While generic algorithm based optimization is powerful, it requires a significant amount of finite element analysis (FEA) calculations. To reduce the heavy calculation, the surrogate model assisted optimization in which the FEA calculations are replaced with surrogate models is drawing attention. There are several options for incorporating surrogate models in the optimization process. Two typical configurations are taken up. One is the configuration in which surrogate model generations are substituted between FEA generations. The other configuration uses surrogate models to expand the exploration scope of each FEA generation. The results show that both configurations have higher cost performance than the original FEA optimization, and the acceleration effects of the two different configurations are at the same level despite the different intentions.

An effective method for training data generation is essential for highly accurate magnetization estimation of permanent magnet using deep neural network (DNN). However, the previous method of generation of training data tended to decrease in accuracy as the number of poles in a multipole magnet is increased, although it could accurately estimate 3D magnetization distribution sequentially. In this paper, we proposed a method for training data generation that combines the Heisenberg model with a 1-D Fourier series expansion. As a result of magnetization estimation using a DNN with this proposed method, it was confirmed that magnetization estimation is highly accuracy for multipole orientation magnet.

This paper presents a method for estimating changes in electromagnetic wave propagation in the 28 GHz band due to scatterers using machine learning. In 28 GHz and mmWave band communications, it is necessary to understand the effects of moving scatterers on electromagnetic wave propagation to dynamically control the radio wave environment. In this research, a convolutional neural network (CNN) is used to quickly understand radio propagation considering scatterers. The analytical model and results of the ray tracing method are used as explanatory and objective variables for training data. The estimation of radio propagation by the CNN is found to be in good agreement with the results of the ray tracing method, with a coefficient of determination of 0.84. Furthermore, it is found that radio propagation with a modified radiation pattern of the transmitting antenna can be estimated with less training data through transfer learning.

In this study, a new optimization method for topology optimization (TO) of permanent magnet synchronous motors (PMSMs) using reinforcement learning is proposed. The method employs the twin delayed deep deterministic policy gradient algorithm, a reinforcement learning approach, for TO. It achieves an 18.7% reduction in computational cost while maintaining search performance compared to the conventional optimization method, under the optimization problem of maximizing average torque and minimizing torque ripple for PMSMs.

This paper proposes a new approach for topology optimization of synchronous motors using Bayesian optimization and a variational auto encoder (VAE). The VAE compress the search space into a latent space, where Bayesian optimization is applied to explore more efficiently. The proposed method obtains a solution 52.3% faster than conventional methods.

The physics-informed neural network (PINN) method is of great interest for a wide range of applications in physics and engineering. Recently, the PINN method based on the Helmholtz equation has been successfully applied to the electromagnetic modeling of power-bus structures on printed circuit board. In the present study, a PINN method based on Maxwell’s equations in the frequency domain is presented as one more possibility for its modeling. This method is verified in comparison with the analytical solution of circular power-bus structure. It is also applied to a rectangular structure with round corners and a voltage distribution is discussed.

A novel multi-objective optimization approach for the magnetic design of power inductors is proposed based on the deep reinforcement learning (DRL). The optimization problem is modeled using an artificial neural network (ANN) with the encoderdecoder architecture. The encoder processes the design specification, while the encoder indirectly generates the optimal design variables by modeling their probability distributions. Trained via reinforcement learning, the design task can be addressed by the DRL-based method without iterative processes, drastically cutting down computational time. Additionally, the advantages in the convergence and diversity of the proposed approach are demonstrated through a case study on a power inductor for inverters.

In motor design, shape optimization aimed at improving performance constitutes a critical issue. Finite Element Analysis (FEA) is widely used for shape optimization due to the high levels of accuracy it provides, but it can lead to long computation times, thereby reducing the overall efficiency of the design process. To address this issue, recent research has focused on the use of surrogate models to quickly estimate motor characteristics. In this paper, we verified a method for estimating the current-torque characteristics of a motor by estimating the coefficients of an approximation curve using a surrogate model and then calculating the torque. As a result, it was found that estimating the coefficients of the approximation curve resulted in worse levels of estimation accuracy when compared to cases where the torque for each current value was estimated.