Glass transition temperature prediction in lignin polyurethanes using machine learning on small experimental dataset

Lignin-based polyurethanes (PUs) offer a compelling route towards sustainable material development, yet the challenge of designing chemical formulations with targeted properties, such as glass transition temperature (Tg), remains unresolved. In this work, we present a systematic approach to explore key structural parameters – such as lignin content, co-polyol chain length, isocyanate functionality, and mixing ratios – across 136 unique formulations, creating a diverse dataset of lignin-based PUs. By harnessing this small dataset, we develop a machine learning (ML) stacking ensemble model capable of reliably predicting Tg, with a mean absolute error of 13.41 °C on the validation set, surpassing the performance of all individual models. Additionally, we enhance model interpretability by integrating advanced mapping techniques and employ an adaptive grid search algorithm to explore extrapolative scenarios. Our workflow, paired with a user-friendly interface, enables rapid discovery and optimization of formulations with desired properties. This study not only deepens the understanding of structure–property relationships in lignin-PUs but also provides a scalable ML-driven tool for designing sustainable materials with precision, highlighting the transformative potential of artificial intelligence in green chemistry and materials innovation.