Benchmarking Neural Architectures for Predicting Properties in Hard-Magnetic Soft Metamaterials

Active metamaterials are materials whose mechanical properties can be tuned by external stimuli, and they have attracted growing research interest in recent years. Yet predicting their behavior remains difficult. The interaction between magnetic and mechanical fields gives rise to nonlinearities that are hard to describe analytically, while experimental data are often scarce and costly to obtain. In this paper, we compare three neural network architectures, namely MLP, PINN, and KAN, for predicting Properties in Hard-Magnetic Soft Metamaterials, using a synthetic dataset constructed from physical principles and calibrated against experimental bounds (availaible at https://www.kaggle.com/datasets/haianhjobs/magneto-dataset). The dataset is designed to reflect the key characteristics of the material, including magnetic softening, auxetic phase transitions, and the angular dependence of the mechanical response. All three models yield acceptable predictive results. The most meaningful differences, however, lie not in accuracy but in generalization and parameter efficiency. More notably, KAN reconstructs the angular relationship in the data through symbolic regression without this structure being explicitly encoded, something neither MLP nor PINN is able to do.