Artificial neural network modeling for prediction and optimization of biochar yield and properties

Biochar produced via biomass pyrolysis offers promise for soil amendment, carbon sequestration and renewable energy applications. This study presents an Artificial Neural Network framework (ANN) coupled with Particle Swarm Optimization (PSO) for accurate prediction and optimization of biochar yield, higher heating value (HHV₂) and carbon content(C₂). This work addresses the gap in multi-objective biochar design by integrating predictive modeling, global and local explainability, and optimization into a unified framework. A dataset of 296 experimental runs, covering pyrolysis temperatures(θ), residence times(t), elemental composition, proximate analysis, and initial heating value(HHV₁), was normalized, validated and split into training-validation (80:20) subsets. Six feedforward ANN models with varied input combinations were tuned via a two-stage grid and randomized search, achieving an overall R² of 0.909 and average RMSE of 3.15 across outputs. A further 5‑fold cross‑validation on the selected Model 1 (with 11 inputs) yielded mean ± std dev R² of 0.895 ± 0.013 and RMSE of 5.71 ± 0.31 % for yield, confirming the model’s robustness. However, model 4 with the lowest inputs (θ, t, and HHV₁) also predicted an appreciable average R² of 0.870 and RMSE of 3.84. Feature-importance, partial-dependence and SHAP analyses identified pyrolysis temperature and volatile matter as primary drivers of biochar properties. PSO yielded global optimum conditions at 200°C and 47 min (69.3 % yield, 16.9MJ/kg HHV₂, 45.2 % C₂), with tailored settings for agricultural residues (509°C-52min) and woody biomass (405°C-85min) to balance energy density and yield. These results demonstrate a robust, data-driven approach for designing biochar production processes, with potential for real-time control and multi-objective optimization in future applications.