Conference Agenda

The magnetic field distribution in the armature-rail system is a critical parameter for evaluating the performance of electromagnetic propulsion device, because it directly impacts the armature motion and energy conversion efficiency. Machine learning has emerged as an effective approach for predicting the magnetic field, enabling real-time simulation of the armature-rail system and rapid calculations in digital twin scenarios. However, purely data-driven neural network models lack interpretability of physical mechanisms and struggle to achieve accurate predictions with limited training samples. Therefore, this paper proposes a dual-driven model prediction method for the magnetic field distribution in armature-rail system of electromagnetic propulsion device by introducing physical constraints. Firstly, based on the non-periodic pulsed excitation condition, physical constraints of the magnetic diffusion equation are established in the frequency domain and integrated as regularization terms into the neural network model for magnetic field prediction. Next, the magnetic vector potential in the armature-rail region under varying excitation currents and conductivities are numerical simulated to construct the training dataset. Spatiotemporal data is enabled to realize the training under the coupling of network model and frequency domain control constraints. Finally, the feasibility and prediction accuracy of the model are validated using a test set. The work in this paper provides theoretical support and a valuable reference for the further reliability prediction and optimal design of electromagnetic propulsion device.

The launch efficiency of the multistage synchronous induction coilgun (MSSICG) is highly dependent on the trigger positions. However, due to the complex nonlinear characteristics and the dynamic coupling of the MSSICG, identifying the best trigger position is a challenging task, particularly when different mass of loads are applied. This article proposes a hybrid model that combines both data-driven and physics-driven approaches for efficiently determining the best trigger positions. A deep neural network (DNN) is employed to predict the best trigger position for a single-stage launcher. The outputs of the DNN serve as intermediate variables, which are then fed into the physics-driven model to compute the sequence of best trigger positions for the MSSICG under different load conditions. Finally, a three-stage induction coilgun is used to validate the accuracy and effectiveness of the proposed hybrid model.

To improve the computational efficiency of the magnetic field prediction of permanent magnet (PM) motors, an improved numerical U-net neural network model based on linear activation function is proposed in this paper. First, to address the pixel-level error and poor generalization under the pixel prediction of traditional U-net network, the output linear activation function is utilized to combine with the U-net network. And a two-dimensional array of magnetic field distributions is derived to achieve the numerical prediction. Secondly, a parallel convolutional layer is utilized to import two-channel inputs, representing the geometry and excitation dimensions, respectively. Finally, the magnetic field distribution of the motor under different excitations with different geometries and rotation angles are predicted. By comparing with the results of finite elements method (FEM), the improved numerical U-net neural network is 13 times more efficient in the range of peak motor excitation currents from 50 A to 250 A, and the average percentage error is within 4%.

Pulsed high magnetic fields (PHMFs) have emerged as essential tools in scientific research. As the key to generating pulsed high magnetic field, the design and analysis of pulsed magnets has become the focus of scientific instrument research and development. Design methodologies for pulsed magnets involve a multidisciplinary approach, in which high-frequency circuit models are critical to the generation of magnetic fields. The model can be used to simulate the electromagnetic behavior of pulsed magnets and predict the discharge process. In this paper, we propose a high-frequency modeling method based on neural networks, which is developed through impedance measurements and neural network curve fitting. The impedance of the pulsed magnet was measured using a designed fixture adapter and a vector network analyzer (VNA), and high-frequency modeling and element value prediction were performed through a neural network. The method has high precision, good versatility and simple parameterization

Metamaterials are artificially structured materials with unique properties not found in nature, offering significant potential in manipulating electromagnetic fields. However, when applying numerical methods to solve metamaterial-related problems, challenges arise in balancing the computational efficiency and the accuracy. A physics-informed neural network based method is thus proposed in this paper to address the low-frequency electromagnetic fields including metamaterials in the time domain. The proposed method demonstrates a comparable accuracy to the time-domain finite-difference method.

Magnetic field reconstruction is a critical task with wide applications in design optimization, inverse problems, and other fields requiring precise electromagnetic field modeling. This paper proposes a novel conditional variational autoencoder-based model that demonstrates significant potential in reconstructing magnetic fields for random radius distributions in the TEAM benchmark problem 35 without relying on prior field conditions. The method utilizes coil distribution images as conditional labels and field distributions as the reconstruction targets, achieving a robust convergence and a low reconstruction error. Experimental results on randomly generated test cases in TEAM problem 35, with arbitrary radius distributions, show that the proposed approach accurately reconstructs fields.

The finite element modeling of hysteresis and eddy currents in ferromagnetic laminated cores is intricate and costly. In consequence, magnetic losses are often only evaluated a posteriori in electrical machine modeling. This leads however to potentially inaccurate results, as the effect of magnetic losses on the field computation is neglected. Deep learning offers an efficient solution to this problem. Two neural network architectures able to replicate hysteretic behaviors are discussed in this paper. Their fast and accurate inference makes them good candidates to serve as material laws in multiscale finite element simulations.

This paper explores the application of Physics- Informed Neural Networks (PINNs) for addressing a magnetostatic problem with particular attention to material interfaces. A representative test case featuring an inductor composed of two different materials is employed to solve Poisson’s equation. The results generated by the PINN are compared to those obtained using the Finite Element Method (FEM).

For high-dimensional parametric problems, a data-based surrogate model can be generated from simulation or experimental data to quickly evaluate quantities of interest, which can then be applied in optimization or real-time monitoring or control scenarios. Deep neural network (DNN) has become an important approach for surrogate modeling of high-dimensional parametric problems. However, a large amount of data is a prerequisite for training an DNN model. Numerical simulation is an important method for data acquisition. For an DNN model trained by simulation data, an essential problem is how to obtain sufficient data with relatively low computational cost. The fidelity and efficiency of numerical simulation are highly related to the density of meshes. In this paper, we propose an incremental learning (IL) strategy combined with simulation data of different fidelity, which are calculated with meshes of different density, for the efficient training of the DNN-based surrogate model. The proposed method is validated by an example for construction of IC parasitic capacitance pattern library. The results show that a surrogate model with the same accuracy can be developed with lower computational cost from the aspect of data preparation with the presented strategy compared to the traditional DNN.

Current calculation methods for interturn short-circuit (ITSC) faults face significant challenges in achieving an optimal balance between computational accuracy and efficiency. An efficient calculation method for ITSC faults is developed in this work using a fuzzy neural network (FNN) surrogate model. The finite element method (FEM) model of transformer ITSC faults is used to generate training, validation, and testing datasets. Different inputs, including the number of shorted turns and the fault position, are investigated to determine the leakage factors. These leakage factors are then integrated into the circuit-based model to obtain the calculated results. The proposed method effectively captures the variation in leakage factors under different conditions, offering a comprehensive combination of data-driven and model-based approaches. Comparisons between the surrogate model results and the field-circuit simulation results reveal that the proposed method is calculated accurately and more rapidly.

As an important research direction in the fields of science and engineering, the electromagnetic inverse problem is with the core goal of inferring the internal physical parameters of the source or system by observing the external response of the electromagnetic field (such as the distribution of the electric field or magnetic field). Due to the lack of conventional stability, nonlinearity and non-convexity, it is quite challenging and difficult to solve these problems with traditional numerical methods. With the development of computational power, an increasing number of methods based on Scientific Machine Learning (SciML) has been utilized in solving forward and inverse problems, like Physics-Informed Neural Network (PINN), Deep Ritz Method (DRM) and Weak Adversarial networks (WAN). Nevertheless, the physical governing equation employed by these inverse problems are defined within the solution domain. This necessitates the measurement data from internal domain for the inverse prediction of unknown model parameters and the forward computation of unknown physical field quantities. To overcome such problems, we have already proposed a new network based on boundary information to solve inverse problem called Boundary Informed Alone Network (BIAN), which can predict the unknown distribution with only two kinds of boundary conditions. However, for most electromagnetic inverse problems, the goal is not to determine the medium distribution across the entire domain, which is often partially known or predefined. Instead, the focus is on reconstructing the unknown medium distribution within a specific subregion of interest. In this paper, we proposed an adaptive regularization method to develop the behavior of BIAN in dealing with piecewise distribution problems. By introducing adaptive regularization, the solution accuracy of BIAN has been significantly improved, particularly near the medium interfaces.

This paper introduces a novel approach for predicting the scattering parameters (S-parameters) of metasurfaces using a deep learning model based on point clouds. The proposed methodology involves encoding both the spatial and physical characteristics of metasurfaces into point clouds, which are then processed by a proposed deep learning model. The model architecture includes multi-scale pooling layers to retain critical frequency-domain information and dense layers for precise regression. By leveraging a training dataset of 1024 samples and employing efficient feature extraction techniques, the model successfully predicts S-parameters within the 2–12 GHz frequency range of metasurfaces. Experimental results validate the model’s high accuracy, with predictions closely aligning with those obtained from full-wave simulations. This work highlights the potential of deep learning to expedite metasurface design and optimization, offering a computationally efficient alternative to traditional numerical methods.