Interpretable Machine Learning of Hydrogen Production from Reactive Molecular Dynamics Simulations of Polyethylene Oxidative Gasification

Reactive molecular dynamics (ReaxFF MD) simulations of polyethylene (PE) oxidative gasification generate large trajectory datasets suited to machine learning (ML). We recast H2 production monitoring as a leakage-free forecasting task: given the simulated thermochemical state at time t, predict the time-weighted cumulative H2 production cumH2(t+h) a fixed horizon h=25 ps ahead. Data from 48 trajectories spanning six O2 loadings (1–900 molecules per 50 PE chains) and eight temperatures (1500–2900 K) yield 72,117 forecast pairs (45 simulations; three truncated trajectories excluded). Eight tabular regression models are benchmarked against persistence and physics-informed baselines under simulation-grouped nested cross-validation, a held-out test partition, temperature- and loading-holdout stress tests, and a paired ablation removing all H2-derived inputs. With the H2 history available, the best test model (random forest) attains R2=0.9999 at RMSE 121.8 molecules·ps, a 93.0% RMSE reduction relative to persistence, robust to unseen O2 loadings and largely robust to unseen temperatures. Without it, no model beats persistence in any evaluation: the best ML model (an artificial neural network) is 26.0% worse, and accuracy decreases substantially under temperature extrapolation. The ablation attributes 93.3–98.3% of the tree and kernel models’ RMSE reduction to the H2-derived features, quantifying how much forecasting skill derives from the target history rather than instantaneous observables.