Abstract

Data-driven interatomic potentials have become an indispensable tool in the approximation of the potential energy surface with quantum mechanical accuracy. One of the latest frameworks, Atomic Cluster Expansion (ACE), has reduced evaluation times to 0.1 ms/atom/core, enabling researchers to investigate length and time scales previously unfeasible while respecting the fundamental laws of nature. The ACE framework also allows for excellent interpretability of the potential and allows modellers to carefully study the relevance of body-ordered interactions in quantum many-body atomic systems. Central to the development of interatomic potentials is the generation of the training database. In order to aid this task Hyperactive Learning (HAL) has been developed to rapidly build ACE potentials from scratch. HAL generates valuable training configurations by adding a biasing term towards uncertainty, in turn accelerating database generation by up to an order of magnitude compared to standard AL. General HAL protocols for building ACE potentials for polymers and alloys will be presented and applications such as determining polymer density and predicting alloy phase transitions discussed. These studies will give insight into physical phenomena such as precipitate formation and chemical ordering in alloys.