A Physics-Informed Neural Network Framework for Modeling and Optimization of Biomass–Plastic Co-Pyrolysis Systems

Biomass–plastic co-pyrolysis is a promising route for sustainable fuel production; however, predictive modeling is limited by the inability of conventional kinetics to capture composition-dependent interactions and by thermodynamic inconsistencies in purely data-driven models. This study develops a physics-informed neural network (PINN) that integrates composition-dependent Arrhenius kinetics with global mass and elemental (C, H, O) conservation through a multi-objective loss formulation. The model was trained on 106 experimental datasets and achieved high predictive accuracy on unseen data (R2 = 0.988, RMSE = 1.838 wt%, MAPE = 2.4%), corresponding to a 66% reduction in RMSE relative to an XGBoost baseline. The PINN strictly preserved mass balance (0.00% violation) and demonstrated stable extrapolation up to 1000 °C. External validation confirmed strong generalization (R2 = 0.991, RMSE = 1.74 wt%) with mass closure error below 0.5 wt%. Beyond prediction, the framework enabled mechanistic interpretation by identifying a critical compositional threshold at 30 wt% plastic, consistent with the onset of hydrogen-transfer interactions reported in the literature, reducing the effective activation energy from 180 to 85 kJ/mol and enhancing bio-oil yield by up to 11.9 wt% above additive behavior. The trained PINN was further deployed as a differentiable surrogate for gradient-based optimization, identifying optimal conditions near 513 °C and high plastic fractions (> 90 wt%), with predicted yields exceeding 92 wt%. The results demonstrate that physics-informed learning enables thermodynamically consistent, interpretable, and efficient modeling, establishing a domain-aware predictive surrogate for laboratory-scale co-pyrolysis process optimization.