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

Area of Science:

• Computational Chemistry
• Materials Science
• Environmental Science

Background:

• Accurate modeling of gas/solid interfaces is crucial for environmental applications like sensing.
• Developing classical force fields (FFs) with quantum-level accuracy for complex systems remains a challenge.
• Integrating active learning (AL) with physics-aware potentials offers a promising solution.

Purpose of the Study:

• To propose an active learning framework for developing classical force fields (FFs).
• To accurately model potential energy surfaces (PES) of gas/solid atomic-scale complexes.
• To achieve quantum-level accuracy for interfacial systems using efficient, physics-aware potentials.

Main Methods:

• Trained physics-aware potentials on actively sampled density functional theory (DFT) data.
• Utilized adaptable semiempirical descriptors optimized via Pareto analysis.
• Generated candidate structures using Metropolis Hastings Monte Carlo (MHMC) or stochastic molecular dynamics (sMD).
• Selected DFT candidates using an outlier score (OS) for diverse PES exploration.

Main Results:

• Developed FFs capable of capturing cohesive, physisorption, and chemisorption interactions with high accuracy.
• Achieved accuracy close to ab initio methods while retaining semiempirical potential efficiency.
• Demonstrated FF utility in molecular dynamics (MD) simulations of silver clusters in gas phases.

Conclusions:

• The proposed AL framework enables the development of accurate and efficient classical force fields.
• The methodology is versatile and adaptable to various descriptors, basis sets, and sampling techniques.
• This approach advances the accurate modeling of atomic-scale interactions for environmental applications.