Wood Science and Technology Промышленная и производственная инженерия

Deciphering wood pyrolysis dynamics: SHAP-enhanced machine learning models for biochar optimisation

Özbay, Günay, Goksal, Fatma Pinar
Springer Nature · 2026-09-22 · DOI 10.1007/s00226-026-01809-z

Аннотация

Biochar derived from biomass pyrolysis is an essential resource for carbon sequestration and renewable energy generation, but the complex, nonlinear relationships between feedstock biochemistry and operational variables often hinder the efficiency of conventional experimental methods. This study aims to investigate the thermochemical dynamics of wood pyrolysis using a high-fidelity “Digital Twin” framework integrating Artificial Neural Networks (ANNs) and SHapley Additive exPlanations (SHAP). Based on primary experimental data from Pinus sylvestris L. and Fagus orientalis Lipsky, a Multilayer Perceptron (MLP) model was developed with remarkable prediction accuracy for biochar yield (R^2 = 0.987) and carbon content (R^2 = 0.991). In addition to the statistical results, SHAP analysis was also used to interpret the model’s “black box” logic in a mechanistic manner. As demonstrated, while pyrolysis temperature is the most important factor in reducing yield due to the volatile phase transition, lignin content contributes to a structural resistance mechanism that stabilises the carbon skeleton, particularly during intense thermal decomposition at approximately 500 °C. These findings, corroborated with recent literature (2022–2025), indicate that the ANN-SHAP coupling is a suitable compromise between machine learning and chemical kinetics. The paper presents a scalable, explainable AI pathway for evaluating compositional and operational interactions in woody biomass pyrolysis.

Читать у издателя →