Machine learning prediction of bio-oil yield from biomass and plastic co-pyrolysis based on condensation conditions

This study develops and compares Deep Neural Networks (DNN), Lightweight Gradient Boosting Machine (LightGBM), and Support Vector Machine (SVM) to predict bio-oil yield from biomass – plastic co-pyrolysis. A dataset of 152 experimental samples was constructed from published literature, incorporating feedstock properties, pyrolysis conditions, plastic ratio, and, for the first time, condensation parameters (condensation temperature, stages, and coolant type). Results indicate that introducing condensation conditions significantly enhances the predictive performance of DNN and LightGBM models (p < .05). For DNN, the test R2 increased from 0.897 to 0.917, with MAE and RMSE reduced to 3.987 and 5.278, respectively. SHAP analysis reveals that plastic ratio remains the dominant factor, while condensation parameters substantially influence feature importance and liquid recovery efficiency. These findings demonstrate that incorporating condensation conditions improves both prediction accuracy and process interpretability, providing a more comprehensive framework for bio-oil yield optimization.