Real-time process design enabled by an interpretable GBDT model for high-fidelity prediction of pyrolysis products from crop residues.

The precise prediction of product distribution during crop biomass pyrolysis is critical for its valorization but remains challenging due to complex nonlinear interactions between feedstock heterogeneity and process parameters. This study developed an interpretable machine learning (ML) framework integrating Gradient Boosting Decision Trees (GBDT) optimized via Bayesian hyperparameter tuning (Optuna) to elucidate the coupling effects of proximate/ultimate composition and temperature on biochar, bio-oil, and gas yields. The GBDT model, trained on 265 data points from 28 types of crop residues, demonstrated the highest accuracy (test set R2 ≥ 0.89). Mechanistic interpretation using Shapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) identified the O/C molar ratio as the most critical feature governing product distribution, with a threshold (O/C > 0.4) triggering pronounced deoxygenation towards gases. Ash content catalytically promoted secondary cracking reactions via inherent alkali metal oxides (e.g., K2O), reducing bio-oil yield, while pyrolysis temperature governed the selection of reaction pathways. Through the innovative integration of ML with Response Surface Methodology (RSM), distinct optimal process windows were established for maximizing each product: biochar (300 °C, 12 °C/min), bio-oil (523 °C, 79 °C/min), and gas (713 °C, 103 °C/min). Furthermore, an online predictive tool was developed to enable intelligent real-time regulation, effectively bridging the gap between model-based insight and industrial implementation of optimized pyrolysis processes.