SCALABLE MACHINE LEARNING FOR TEMPERATURE-DEPENDENT CONVERSIONS IN COBALT-CATALYZED METHANE DRY REFORMING

Abstract

Predictions of methane dry reforming (DRM) over cobalt catalysts through machine learning have been limited by data sparsity, and the original 57-point benchmark has limited test accuracy (R²=0.42-0.67) because predictions and experimental measurements are overfitting on limited experimental measurements. The research generates a 402-point temperature-dependent dataset (550-750°C) that eliminates this fundamental tradeoff, providing ensemble forecasts with unprecedented R²=0.971 (CH₄ conversion) and R²=0.958 (CO₂ conversion), which are 70 and 80% better, respectively. Polynomial baseline gradient boosting (XGBoost/LightGBM) models outperform neural architectures by direct control over overfitting ( -2<0.055 vs 0.33-0.48), and SHAP analysis proves kinetic dominance of temperature (98.7% kinetic) allows single-sensor control of processes, compared to the complexity of multi-variable neural networks previously, while acceptable consistency of 10-fold cross-validation ( -2≈0.008) and production scale stability. The temperature-only model yields Arrhenius activation, kinetic acceleration, and equilibrium plateau with less than 2% accuracy, measuring the systematic 5-10% lag of CO₂ in greenhouse gas fuels to achieve the desired yield plateau. This gracefully adequate design methodology makes frontier performances with 1/7^th of the current literature 300–1000-point data requirements and multidimensional instrumentation requirements to deploy a digital twin of DRM, where research percent conversion accuracy translates to 0.15/kg of syngas cost reduction under real-world thermal gradient conditions of ±50^oC. The open-source pipelines democratize entry, where the catalyst screening is guided, whilst the scale of datasets is defined to be the final choice in reproducible catalytic forecasting.