Hybrid evolutionary machine learning framework optimizing biochar production in biomass pyrolysis

Growing interest in renewable energy worldwide has heightened the need for efficient modeling of biomass conversion. This study proposes a hybrid framework coupling five machine-learning models (ELM, ANN, RF, SVR, and XGB) with six population-based optimizers (SGA, PSO, DE, ABC, GWO, and xNES) to predict biochar yield in biomass pyrolysis. The database contains 423 experimental records from 44 lignocellulosic feedstocks, integrating proximate/elemental properties with operating conditions. Five learning algorithms were benchmarked, and each was tuned by every optimizer, enabling a systematic direct comparison of 30 hybrid configurations. Evolutionary hyperparameter tuning, evaluated using 5-fold cross-validation and 30 stratified train–test repetitions, improved accuracy and generalization relative to non-optimized learners. The best configuration (ABC–XGB) achieved R 2 = 0 . 815 , RMSE = 3.845, and MAE = 2.199 on the test sets. ANOVA ( p < 0 . 05 ) indicates that performance differences are driven mainly by the ML estimator rather than the optimizer, although DE, ABC, and GWO provided stable tuning across runs. Reliability was further examined using propagation-based uncertainty quantification with bounded input sampling, yielding narrow confidence intervals, near-zero mean prediction error, and low variability. The proposed evolutionary-ML surrogate model captures nonlinear thermochemical relationships and supports the design, optimization, and control of biomass pyrolysis and related bioenergy systems. • Hybrid ML-evolutionary framework predicts biochar yield in biomass pyrolysis. • Benchmarked 5 ML models with 6 metaheuristics on 423 experiments (44 biomasses). • Evolutionary tuning improved accuracy and generalization versus untuned models. • Best model (ABC-XGBoost) reached R 2 =0.815, RMSE=3.845, MAE=2.199 on test data. • Uncertainty quantification showed low bias and narrow prediction intervals overall.