A systematic hybrid mechanistic–machine learning framework for catalytic reactor modelling and computational validation using CO oxidation

Accurately forecasting the fast transients that govern catalytic reactors remains difficult because first-principles ordinary differential equation (ODE) models neglect unmodelled heat and mass-transfer effects and therefore perform poorly (baseline CO-oxidation rate R² = –0.231). For the above reason, this study presents a systematic hybrid mechanistic machine-learning (ML) framework that couples a physically rigorous CSTR model with data-driven residual learning to close these physics gaps. A six-factor design of experiments generated 500 operating scenarios, and after simulation, quality screening, derivative estimation, and residual/outlier filtering, the residual-learning dataset comprised approximately 33,096 usable samples. Five regressors (XGBoost, LightGBM, SVR, MLP and sparse Gaussian-process regression) were hyperparameter-tuned with Optuna and blended through weight optimisation. Uncertainty was propagated with GP posterior bands and inter-model disagreement. The optimised ensemble lifted test-set accuracy to R² = 0.755, RMSE = 0.006 mol·m⁻³·s⁻¹ and MdAPE = 93 % a dramatic recovery over the mechanistic baseline. ±2σ GP bands captured 94 % of unseen points, providing actionable epistemic bounds. Performance deteriorated by only ∼21 % when 5 % Gaussian sensor noise was injected, confirming robustness for on-line use. By modularising experiment design, physics-guided feature engineering, automated model selection, and calibrated uncertainty quantification, this workflow delivers interpretable, real-time-capable surrogate models within the modelled operating envelope, outperforming pure ODE and single-model ML baselines. The protocol is transferable to other catalytic systems and establishes a reproducible path toward uncertainty-aware reactor optimisation and control.