Physics-Informed Neural Anhysteresis Surrogate for Magneto-Elastic Vector Hysteresis in Device Simulations

In many applications, accurately capturing the magneto-mechanical coupling and dissipative effects at the material level is essential for realistic simulations. Embedding the Simplified Multiscale Model (SMSM) inside an energy-based hysteresis framework yields high fidelity but is computationally intensive for three dimensional finite element (FE) analyses. This paper introduces NNSMSM, a physics-informed multi-task deep neural network that emulates the expensive SMSM operator. A hybrid Latin-hypercube/Sobol sampling strategy efficiently explores the magneto-mechanical loading space. The network is trained with a composite loss that simultaneously fits magnetisation and magnetostrictive strain while enforcing reciprocity and positive definiteness of the susceptibility tensor. The traced TorchScript model is linked to the open source FE software openCFS, replacing the SMSM inside the Vector Play Model (VPM) hysteresis model with zero code changes. The benchmark of a PMSM device simulation shows a speed-up of wall clock time by a factor of 11 while preserving global accuracy of hysteresis losses.