Hessian-based molecular conformation augmentation for a scalable and efficient strategy of machine learning interatomic potentials

The researchers proposed a molecular conformation enhancement strategy based on the Hessian matrix, aimed at improving the scalability and efficiency of atomic potential functions in machine learning. The method includes two data enhancement schemes: Unitary Gaussian displacement (UniAug) and mode-weighted displacement (ModeAug). Effective enhancement is achieved through simple Taylor expansion, without modifying the training objective or expanding the automatic differential graph, allowing for seamless integration into existing architectures and training processes. Comprehensive evaluations on balanced and unbalanced datasets showed that this strategy not only improved model accuracy but also provided practical guidance for specific tasks.