Active-learning strategy for the development of application-specific machine-learning force fields

Emerging data-driven approaches in materials science have triggered the development of a plethora of machine-learning (ML) force fields (FFs). In practice, they are constructed by training a statistical model on a reference database to predict potential energy or atomic forces. While most of the FFs can accurately recover the reference data, some of them are becoming useful for actual molecular dynamics simulations. In this work, we develop a simple active-learning strategy for the development of ML FFs targeted at specific simulations (applications). The strategy involves (1) preparing and fingerprinting a diverse reference database of atomic configurations and forces, (2) generating a pool of ML FFs by learning the reference data, (3) validating the FFs against a series of targeted applications, and (4) selectively and recursively improving the FFs that remain unsuitable for a given application while keeping their performance on other applications uncompromised. We demonstrate this strategy by developing a series Al and Cu ML FFs that can simultaneously be used for various applications, including (elastic) stress/strain analysis, stacking-fault energy calculations, and melting simulations. This strategy is generalizable, i.e., it may be used for other materials as well.