Insights into the capability of machine learning for predicting carbon flow in biowaste hydrothermal treatment under non-catalytic conditions

Hydrothermal technology can efficiently and highly valorize biomass waste into solid phase products (hydrochar), aqueous phase products, oil phase products, and gas phase products. The yield and characteristics of these products are significantly influenced by hydrothermal conditions, leading to cumbersome experiments and high cost inputs. Machine learning can predict the characteristics of target products under new conditions based on existing data. This study aims to predict the hydrothermal carbon flow of biomass waste under non-catalytic conditions. A hydrothermal carbon flow dataset for biowaste under non-catalytic conditions was established, and four machine learning models (ANN, GPR, PSO-LS-SVM, and RF) were used to predict the carbon flow of hydrothermal products. The results showed that the PSO-LS-SVM model has the highest prediction accuracy and stability (The highest R² value is >0.99 with a median >0.92). Feature importance analysis of the optimal model (PSO-LS-SVM) revealed that the biomass waste mass and vessel volume are the most critical parameters for predicting the mass of hydrochar, aqueous phase, and oil phase products, with significantly higher contributions than other variables and positive effects. Validation with new environmental data demonstrated that the model has excellent predictive capability, with minimal deviations between the predicted and experimental values. The R² values were mainly above 0.7, and some reach as high as 0.99. This study provides an important methodological support for the hydrothermal conversion of biomass waste. • A hydrothermal carbon flow dataset for biomass waste was established. • Four models were used to predict the carbon flow of hydrothermal products. • The PSO-LS-SVM model had the highest prediction accuracy and stability. • Feature importance analysis of the PSO-LS-SVM model was conducted. • The PSO-LS-SVM model has excellent predictive capability in new data.