PhyTENG: A physics-informed machine learning framework for modelling contact - separation triboelectric nanogenerators

Triboelectric nanogenerators (TENGs) are promising for harvesting mechanical energy and enabling self-powered sensing. Their electrical performance depends on complex interactions among materials, device architecture, and operating conditions and it is challenging to make accurate prediction models. In this work, we present PhyTENG, a physics-informed machine learning (ML) framework to predict output voltage and current in contact-separation (CS) TENGs. A diverse experimental dataset was curated from literature and augmented with open-source materials databases and ML-based property predictors to obtain atomic- and surface-level physics-based descriptors, including electron affinity, ionisation energy, and triboelectric charge density. Four interpretable gradient-boosting models were then trained to relate these properties and other descriptors to their electrical performance. We further demonstrate the advantages of using physics-inspired descriptors relative to conventional categorical materials featurisation. Overall, the presented PhyTENG framework provides a generalisable predictive tool for TENG optimisation and a pathway to integrate AI-driven materials discovery with AI-driven device design.