機械的に剥離可能な二次元ナノ材料の予測と設計のための物理的に解釈可能な記述子

The mechanical exfoliability of two-dimensional (2D) nanomaterials is fundamentally governed by the exfoliation energy (E_ex). Conventional strategies for determining E_ex mainly depend on either indirect experimental approaches or computationally demanding first-principles calculation, which impose significant demands on time and computational resources. Data-driven methods provide an efficient alternative, yet the limited interpretability of machine learning predictions often restricts their applicability. Herein, we construct a physically interpretable descriptor P₄₃, by coupling a CatBoost model (RMSE = 57.83 meV/atom) with the Sure Independence Screening and Sparsifying Operator (SISSO) for predicting and designing mechanically exfoliable 2D nanomaterials. Remarkably, P₄₃ achieves a recall rate of 99.0% in identifying mechanically exfoliable 2D nanomaterials using four concise parameters. Further analysis of its electronic (φ) and geometric (θ) contributions reveals that φ and θ play dominant roles in predicting low-E_ex (<150 meV/atom) and high-E_ex (>350 meV/atom) 2D nanomaterials, respectively. Leveraging P₄₃ within the element substitution method (ESM) framework, we screened A- and B-site combinations and discovered 8946 novel AB-type 2D structures, including 111 candidates confirmed as exfoliable in the 2DMatPedia database. Overall, this study provides new insights into the E_ex, highlighting the significant role of P₄₃ in accelerating ESM for the development of novel 2D nanomaterials.