Neural Surrogates for Wave-Scattering Inverse Problems Scale with Dynamic Training Examples
Researchers introduce a method to train inductively scalable, single-step neural surrogates for wave-scattering inverse problems. The approach dynamically generates salient training examples via gradient ascent, searching refractive-index and source configurations where the surrogate disagrees with a full-wave ground-truth simulator. Combined with source and ground-truth normalization and an evolving replay dataset, this overcomes the previous limitation of scaling to only tens of simulation variables. The surrogates aim to replace traditional electromagnetic wave simulators like finite-difference time-domain (FDTD), solving forward and inverse problems orders of magnitude faster. The paper is available on arXiv under ID 2608.17344.
Key facts
- The paper is on arXiv with ID 2608.17344.
- Neural network surrogates are an emerging alternative to FDTD simulators.
- The goal is to replace rigorous physical simulations with pre-trained neural networks.
- Nonrecurrent, single-step surrogates previously scaled only to a few tens of simulation variables.
- The new algorithm runs in parallel with surrogate training.
- It uses gradient ascent to search refractive-index and source configurations.
- It identifies cases where the surrogate disagrees with a full-wave ground-truth simulator.
- Source and ground-truth normalization with an evolving replay dataset stabilize and accelerate learning.
Entities
—