Physics-Aware Gradient Boosting Surrogate for Nuclear Fission Rate Prediction

This work presents a physics-aware machine learning surrogate for rapid nuclear fission rate prediction using Gradient Boosting trained on Monte Carlo neutron transport simulation data. The study demonstrates that high global accuracy metrics can conceal severe localized prediction failures in fertile nuclear material regimes, where uncalibrated models over-predict fission rates by factors approaching 2,000× despite near-perfect overall R² scores. To address this, a physics-aware calibration framework is introduced that applies composition-based correction logic without retraining the underlying model. The calibrated surrogate achieves strong agreement with theoretical ground truth while retaining over 1,000× computational speedup compared to direct Monte Carlo simulation. The work combines computational nuclear physics, neutron transport modeling, and machine learning surrogate methods, with applications in rapid nuclear fuel screening and physics-constrained AI systems.