Active Learning of Atomic Size Gas/Solid Potential Energy Surfaces via Physics Aware Models

We propose an active learning (AL) framework to develop force fields (FFs) that accurately model the potential energy surfaces (PES) of gas/solid atomic-scale complexes. A central challenge is integrating AL with flexible, physics-aware potentials to achieve quantum-level accuracy for complex interfacial systems. Our approach trains physics-aware potentials, with incorporated flexibility and smoothness, on actively sampled Density Functional Theory (DFT) data to describe interactions between undercoordinated atomic silver (Ag) clusters and gaseous pollutants (CO$_2$, CO, SO$_2$), relevant for environmental applications like sensing. The AL process follows three stages: (1) FFs are trained using adaptable physics aware potentials of semi-empirical descriptors, optimized via a Pareto analysis scheme; (2) refined FFs generate candidate structures through Metropolis Hastings Monte Carlo (MHMC) or stochastic molecular dynamics (sMD); (3) a subset of candidates is selected for DFT labeling based on an outlier score (OS), which utilizes the existing data descriptor distributions, ensuring diverse PES exploration. This framework produces FFs capable of capturing cohesive, physisorption, and chemisorption interactions with accuracy comparable to \textit{ab initio} methods and advanced machine learning models, while retaining the efficiency of semi-empirical potentials. Our methodology is highly versatile, easily accommodating various choices of descriptors, model basis sets, and sampling techniques.