Moving in the right direction: Emerging ideas in fitting accurate potentials

Abstract The development of accurate interatomic potentials remains a cornerstone challenge in computational materials science. This article examines the transformative shift from physics-based potentials to machine learning interatomic potentials, highlighting how emerging methodologies are revolutionizing the field. We discuss how data-driven approaches, incorporating advanced optimization and automated workflows adapted from the machine learning community, enable the creation of powerful potentials that capture complex atomic interactions with unprecedented accuracy and flexibility. These advances allow for dynamic adaptation to diverse chemical environments and competing requirements, moving beyond the limitations of traditional physics-based potentials. Looking forward, we consider the potential of foundation models and the concept of “universal potentials,” envisioning a future where a single, adaptable model framework could accurately model a wide range of materials and phenomena. Graphical abstract