Feature Engineering Enhanced Machine Learning Prediction of Pore Properties in Lignin-Derived Nanoporous Carbon for High-Performance Supercapacitor Applications

Lignin-derived porous carbon (LDPC), prepared through catalytic pyrolysis, has attracted significant attention for its use in energy storage. However, optimizing its nanopore structure is often hindered by the trial-and-error nature of experiments and the limited quality of the available data. In this work, we propose a hybrid machine learning framework that enhances data quality through feature generation, thereby improving the predictive accuracy of pore-related properties. By systematically transforming and combining existing features, we generated 2025 candidate features and identified 30 high-correlation features that were integrated into the original data set. This enriched feature space significantly improved model performance: the final regressor achieved R 2 values of 0.99 for specific surface area (SSA) and total pore volume (TPV) and demonstrated strong generalization capability ( R 2 > 0.85). Interpretability analysis further confirmed that the final regressor captures meaningful physical mechanisms in the generated feature space. Our approach provides an efficient pathway to enhance data set quality and accelerate the discovery of high-performance nanoporous carbon materials from lignin. By emphasizing nanoscale pore characteristics, this work highlights the potential of LDPC for high-performance supercapacitor applications.