AI and the World's Hard Problems · Astronomy & space · arXiv · 2026-10-21
A Radiation-Hardened, Fault-Aware Quantized Neural Network That Self-Heals
Deep learning models deployed in radiation-challenging environments such as spacecraft and particle accelerators degrade as high-energy particles corrupt data, motivating a runtime method to repair quantized neural network weights after bit flips. The method, called Projection Healing, extends Fault-Aware Training with an alignment penalty that concentrates undamaged activations and weights into a low-rank subspace, from which flagged corrupted coordinates are reconstructed on commodity hardware without storing extra replicas. Across six regression benchmarks it extended the operational scrub interval by a median of 2.04× at +5% RMSE degradation (versus 1.62× for classical SECDED), and holding the FAT checkpoint fixed while toggling only the runtime heal gave a further ~23% reduction in the native-unit excess integral.