Hybrid Physics-Informed Reinforcement Learning for Plasma Characterization in Electrical Wire Explosions

Accurate prediction of discharge observables in electrical wire explosions is challenging due to the inherent nonlinearities in plasma behavior, especially when experimental datasets are limited. This study presents a hybrid physics-informed modeling framework that combines analytical scaling laws with a reinforcement learning-based residual correction strategy to improve prediction accuracy across the charging-voltage range of 9.0–12.9 kV. The analytical baseline, derived from established physical scaling laws, provides normalized estimates for five key observables, including explosion current, peak voltage, square root of peak power, peak pressure, and light intensity. To address localized discrepancies between baseline predictions and experiments, five independent single-step reinforcement learning agents (tabular Q-learning in a contextual-bandit setting) are trained to learn bounded residual corrections in the normalized space, using charging voltage as the sole active state variable. This approach preserves the physical trends of the analytical model while introducing adaptive data-driven refinement. Performance is evaluated using leave-one-out cross-validation (LOOCV), bootstrap resampling, and a restricted high-voltage holdout test for the explosion-current response. Across the evaluated observables, the hybrid framework generally reduces prediction error and improves agreement with the reconstructed experimental trends while maintaining physically interpretable behavior under sparse-data conditions, indicating that constrained residual learning can effectively enhance wire explosion prediction without sacrificing consistency with established physical scaling laws.