Meta-Learning—Efficient Force-Field Parameterization for Accurate Molecular Simulations

Accurate interatomic potentials in molecular dynamics are crucial for predicting the behavior of materials, enabling exploration of the vast materials landscape. This study focuses on the application of meta-learning techniques for force-field parameterization. Traditional approaches to force-field parameterization rely on global or local optimization strategies, often requiring extensive sampling to achieve satisfactory results when the parameter space is large. However, by leveraging meta-learning techniques, we can minimize the need for extensive sampling by acquiring knowledge about the specific regions in the parametric space that align with the desired properties. Meta-learning involves capturing the experience of navigating the parametric space automatically. For example, to capture multiple material properties like local and global structural features, energetics, mechanical and thermal properties, it is important to understand their mapping to the force-field parameter space. In this research, we have successfully applied meta-learning to various force-fields, enabling accurate descriptions of multiple polymorphs and diverse material properties. The results highlight the effectiveness of meta-learning in improving force-field parameterization for enhanced modeling and simulation of molecular systems.