Screening Single-Atom Dopants in GaSe for Gas Adsorption and Sensing via Machine Learning

Abstract

Developing efficient nanoscale materials for toxic gas capture is critical to alleviating environmental crises. In this study, a highly efficient and low-cost nanoscale material design strategy is proposed targeting three typical toxic gases (CO, NO, and NO2). This strategy integrates density functional theory (DFT) calculations with machine learning (ML) methods to enable the robust evaluation of a vast compositional space, encompassing 25 heteroatom-doped GaSe monolayers. Eight mainstream ML models were deployed, facilitating the identification of optimal benchmark models (KRR for adsorption energy and RF for adsorption distance). In conjunction with SHapley Additive exPlanations (SHAP) analysis, this approach enabled the deconstruction of the underlying structure−property relationships. To rigorously verify the framework’s extrapolation capacity, independent out-of-distribution (OOD) blind tests and binding energy evaluations were executed. Subsequently, guided by multidimensional performance evaluations, three representative systems (Fe-, Ag-, and O-doped GaSe) were selected for in-depth electronic, thermodynamic, and kinetic validationsincorporating ab initio molecular dynamics (AIMD) and potential energy surface (PES) scans. Crucially, the results reveal distinct operational viabilities: the Fe-doped system ensures permanent gas sequestration, whereas the O-doped system demonstrates exceptional potential as a room-temperature reversible sensor with an ultra-fast NO2 recovery time of 3.4 × 10−10 s. Aided by artificial intelligence, this study provides a robust paradigm for the rapid screening and inverse design of nanoscale gas capture materials.