Predicting Wave Reflection and Transmission in Heterogeneous Media via Fourier Operator-Based Transformer Modeling

Researchers Zhe Bai and Hans Johansen have introduced a novel machine learning approach to solving complex wave physics problems. Their paper, "Predicting Wave Reflection and Transmission in Heterogeneous Media via Fourier Operator-Based Transformer Modeling," presents an AI surrogate model that approximates solutions to Maxwell's equations in one-dimensional scenarios involving material interfaces. The model was trained on data derived from high-fidelity Finite Volume simulations, incorporating variations in initial conditions and material properties like the speed of light to learn a broad range of wave-material interaction behaviors.

The core innovation lies in the model's architecture, which combines a vision transformer-based framework with Fourier transforms in the latent space. This allows the model to autoregressively learn both physical and frequency embeddings, ensuring the wave number spectra of its predictions closely align with simulation data. Despite the challenge of modeling discontinuities at material interfaces, test results show the ML solution maintains adequate relative errors below 10% for over 75 autoregressive time step rollouts, even with unknown material properties. The prediction errors exhibit approximately linear growth over time, with a sharp but managed increase at the interface itself.

This work, presented at ACDSA 2026, demonstrates a significant step toward using AI as a surrogate for computationally expensive physics simulations. By accurately predicting how electromagnetic waves reflect and transmit at material boundaries, the model provides a faster, data-driven tool for researchers and engineers in fields like optics, acoustics, and material science, where simulating wave interactions is crucial.