Machine Learning Strategies for Predicting and Optimizing Biochar Yield from Diverse Biomass Feedstocks

Biochar production through biomass pyrolysis offers a sustainable approach to reducing reliance on conventional energy sources while mitigating global warming potential. However, identifying the optimal operational parameters, biomass characteristics, and feedstock types remains highly complex. In this study, seven machine learning (ML) models─RT, RBF, MLP, MLR, GPR, ANFIS, and SVM─were developed and applied to a dataset of diverse biomass feedstocks to predict biochar yield. Model performance was evaluated using the same dataset. The Support Vector Machine (SVM) model achieved the highest accuracy with the lowest error (RMSE = 2.61), followed by the Multilayer Perceptron (MLP) and Radial Basis Function (RBF) models. Multiple Linear Regression (MLR), Gaussian Process Regression (GPR), ANFIS, and RT showed lower predictive performance. Using the optimized SVM model, 44 three-dimensional response surface plots were generated to illustrate both individual and interactive effects of feedstock properties and pyrolysis parameters on biochar yield. These plots revealed the complex relationships between variables and emphasized the importance of parameter optimization. Finally, genetic algorithm (GA) optimization indicated that a model-predicted theoretical maximum biochar yield could reach 100% by adjusting feedstock properties (increasing fixed carbon, decreasing volatile matter, and raising ash content) and process conditions (faster heating rate). This surpassed the best experimental result of 95.89% yield obtained from bamboo biomass.