AI designs nuclear experiments with 97.8% similarity via gradient optimization

A team from [institution not specified] has introduced a novel approach to inverse critical experiment design using deep learning and gradient optimization. The method trains a U-Net convolutional encoder-decoder with a multigroup attention pooling layer on OpenMC-calculated sensitivity vectors from grid-based geometries. This differentiable surrogate model enables gradient-based optimization across the full combinatorial design space of material assignments, maximizing the neutronic similarity metric c_k (correlation coefficient of keff bias) — typically requiring c_k ≥ 0.9 for experiment validity.

The technique was demonstrated on the TN-Americas TN-LC transportation cask using HALEU fuel, where existing experimental coverage is limited. Optimization produced geometries achieving c_k scores of 0.97757, 0.81324, and 0.93276 for three target configurations. The multigroup attention pooling outperformed traditional pooling methods and offered interpretable internal attention patterns. This work shows how AI can dramatically accelerate the design of critical experiments needed to validate next-generation reactors and nuclear fuel concepts.