Magnetic field prediction method for electromagnetic propulsion device with frequency-domain magnetic diffusion constraints

Background The magnetic field directly affects armature motion and is a key parameter in evaluating the performance of an electromagnetic propulsion device (EPD). However, traditional numerical methods lack efficiency, and purely data-driven neural networks fail to ensure accuracy with limited training samples. Objective This paper aims to develop a magnetic field prediction method for EPD that enhances both accuracy and physical consistency by incorporating frequency-domain magnetic diffusion constraints into a data-driven framework. Methods Firstly, the non-periodic pulsed excitation current of the electromagnetic EPD is transformed into discrete harmonic components through Fourier analysis. Subsequently, physical residuals derived from the magnetic diffusion equation are formulated in the frequency domain and embedded as soft physical constraints in the form of regularization terms into the loss function of a data-driven neural network, forming a dual-driven architecture. Next, the magnetic vector potential in the armature-rail system (ARS) is numerically simulated under varying excitation currents and conductivities to construct the training dataset, which is then used to train the neural network model with physical constraints. Results Validation results on the test set demonstrate that the dual-driven model aligns better with underlying physical laws and significantly reduces prediction error. Specifically, with a limited number of known samples, the model reduces the average absolute error by 25.11% and 17.24% compared to the purely data-driven model when using sample sizes of P = 100 and P = 200, respectively. Conclusions This paper proposes a dual-driven magnetic field prediction method constrained by frequency-domain magnetic diffusion, which accurately predicts the armature-rail magnetic field distribution under limited samples while ensuring physical consistency. It provides theoretical support and valuable reference for further reliability prediction and optimized design of EPD.