First-principles calculations informing machine learning framework and visualization system for rapid and generalized gas response prediction in black phosphorus sensors

Gas sensors are vital in practical applications, yet the efficient screening of sensing materials remains a formidable challenge. Conventional trial-and-error approaches are costly, single descriptors fail to capture complex interactions, and multi-parameter combinations introduce nonlinearities. To overcome these limitations, we propose a synergistic strategy that integrates first-principles calculations with machine learning (ML) for rapid prediction of gas sensitivity. Using black phosphorus (BP) as a model system, we evaluated its responsiveness to 21 gases by analyzing adsorption-induced electronic and structural changes. Key descriptors extracted from these calculations were used to train six ML models. The extra trees (ET) model demonstrated exceptional robustness, achieving 96% accuracy with minimal deviation in five-fold cross-validation and top-tier performance in F1-score evaluation. Furthermore, analyses of feature importance and SHapley Additive exPlanations (SHAP) identified adsorption energy, p-orbital center, valence band maximum, conduction band minimum, and Fermi level as the dominant descriptors. We also developed a lightweight, Python-based prediction and visualization system. By inputting only these five key features obtained from first-principles calculations, this tool enables real-time assessment of BP's response to various gas molecules. This integrated approach showcases significant potential for predicting material sensing properties and offers valuable theoretical guidance for gas sensor design.