Researchers have developed a new type of neural network surrogate that can solve complex electromagnetic wave scattering problems much faster than traditional methods. The approach uses dynamic training examples and normalization to improve the accuracy and scalability of the surrogates. This allows for robustly accurate simulations with up to 41,772 controllable variables, and generalizes inductively to larger domains without retraining. The team demonstrated the effectiveness of their method on large-scale forward simulations and inverse design problems. --- Why it matters: This work is significant because it enables the development of fast and accurate neural simulators for photonic inverse design and other wave-scattering inverse problems, which can have a major impact on fields like optics and photonics. Engineers and researchers can now use these tools to design complex optical systems with unprecedented speed and accuracy. Source: https://arxiv.org/abs/2608.17344