A Machine-Learning Pipeline for Predicting Methanol Yield in a Double-Pass Plug-Flow Reactor Using a Synthetic Langmuir–Hinshelwood Dataset

Methanol synthesis from syngas (CO/CO₂/H₂) is vital for sustainable chemical production; however, traditional kinetic models hinder rapid reactor optimisation [1]. We present a reproducible machine-learning pipeline to predict methanol yield in a double-pass plug-flow reactor, utilising a synthetic dataset (n = 5,000) derived from a Langmuir–Hinshelwood (L-H) kinetic model [2]. Operating conditions—temperature (473–573 K), pressure (50–150 bar), first-pass residence time (5–20 s), and second-pass residence time (10–30 s)—were sampled, with a 75% yield cap enforcing thermodynamic limits [3]. Exploratory data analysis revealed a strong temperature correlation (), guiding feature engineering (e.g., pressure/temperature ratio) [4]. Model comparison via 5-fold cross-validation on the training folds (before tuning) showed the Random Forest to be the best-performing candidate (mean R² = 0.9851, 95% CI [0.9827–0.9873]) versus linear, ridge and XGBoost baselines, so it was selected for hyperparameter tuning. After GridSearchCV tuning, the finalized RandomForestRegressor (n_estimators=200, max_depth=20, min_samples_split=2) retrained on the training set achieved R² = 0.9872 and RMSE = 0.0060 on the Box–Cox–transformed test targets. On the inverse-transformed (original) yield scale, the test metrics were R² = 0.9897 and RMSE = 0.0139. Parity and residual plots show small, approximately symmetric residuals (mean ≈ 0), consistent with robust generalisation within the sampled domain. Temperature (importance = 0.65) dominated predictions, aligning with Arrhenius kinetics [5]. Deployed via FastAPI and Streamlit, the tool enables real-time predictions with out-of-range warnings, accelerating low-carbon methanol production for green fuels and digital twins. Despite synthetic data limitations (e.g., no catalyst deactivation), future work targets experimental validation and uncertainty quantification (e.g., SHAP, quantile forests) [6]. Open-source artefacts ensure reproducibility [7], bridging AI and chemical engineering for sustainable process design.