Wood Science and Technology Промышленная и производственная инженерия

Explainable deep learning for predicting lignin removal efficiency in deep eutectic solvent-based biomass fractionation

Garg, Diksha, Bansal, Sumit, Bala, Saroj, Singh, Tavisha
Springer Nature · 2026-10-12 · DOI 10.1007/s00226-026-01813-3

Аннотация

Lignocellulosic biomass is an abundant renewable resource with significant potential for sustainable biofuel and biorefinery applications. Efficient lignin removal is a critical step in biomass valorization because it improves cellulose accessibility and enhances downstream conversion efficiency. Recently, deep eutectic solvents (DES) have emerged as promising green solvents for biomass pretreatment due to their biodegradability, low toxicity, tunable properties, and environmentally friendly nature. However, the efficiency of DES-based lignin extraction is governed by complex nonlinear interactions among the solvent composition, biomass characteristics, and pretreatment conditions, making process optimization challenging. In this study, a deep learning–driven framework is proposed to accurately predict lignin removal efficiency and elucidate the underlying mechanisms of DES-based biomass fractionation. A comprehensive dataset comprising 509 experimental conditions was compiled, incorporating biomass composition, DES properties, and pretreatment parameters. A deep neural network (DNN) model was developed using PyTorch and optimized via the Optuna framework to capture nonlinear relationships and interaction effects. The proposed model achieved strong predictive performance on an independent test set dataset, with R² = 0.86 ± 0.01, RMSE = 6.0 ± 0.2, and MAE = 4.8 ± 0.2, demonstrating high accuracy and generalizability. To enhance interpretability, SHapley Additive exPlanations (SHAP) were employed to quantify feature contributions. The analysis revealed that the hydrogen bond donor acidity/basicity (HBD-pKa/pKb) is the most influential factor, while interaction features such as LSR × LogR_0 and the HBD/HBA ratio × LogR_0 significantly influence lignin removal efficiency. These findings highlight that lignin extraction is governed primarily by interaction-driven mechanisms rather than isolated variables. Compared with conventional tree-based models, the proposed DNN framework more effectively captures higher-order dependencies and provides deeper mechanistic insights. Overall, this study demonstrates the potential of integrating deep learning with explainable AI for the rational design and optimization of DES systems. The findings can support the development of efficient and sustainable biomass pretreatment strategies for future biorefinery applications.

Читать у издателя →