Machine-learned interatomic potentials (MLIPs) have emerged as a powerful tool to transport the predictive accuracy of first principles calculations to larger length and timescales. Importantly, MLIPs enable the accurate modeling of chemical reactivity, which is essential for handling complex processes in materials and is beyond the reach of classical potentials [1]. Owing to their success, the field of MLIPs has seen exponential growth and the methodology is advancing very quickly. Keeping up with these developments, while ensuring the carefully groomed quality standards of the community, has become increasingly difficult for both new and experienced users and developers.

Among the many flavors of MLIPs, Gaussian Approximation Potentials (GAP), the Atomic Cluster Expansion (ACE) and its message-passing variant MACE, as well as their software packages are well-known and enjoy large and overlapping user bases. They are also methodologically related, using a spherical harmonics based expansion of chemical environments as a key ingredient. Among these, GAP is the most mature approach, dating back to the infancy of the field.[1] It remains a benchmark method for developing MLIPs and has been used in a plethora of scientific applications. More recent developments in the form of ACE bring the computational cost down significantly,[2] approaching that of classical empirical force fields, while MACE increases the accuracy and data efficiency of GAP even further.[3] All methods are still advancing fast on several fronts, for example on the inclusion of ML-parametrized physics-based models for long-range interactions (van der Waals and electrostatics) or descriptor compression for accelerated simulation, etc. Beyond the methodological capabilities of GAP and (M)ACE, the standardization and automation of their fitting to first principles data has been a strong focus to increase accessibility to users, resulting in software for customizable workflows and quality control.[4]

The third GAP/(M)ACE Developers & Users Meeting aims to bring together the community developing GAP and (M)ACE as well as the user base applying the code in cutting-edge scientific applications. Intensive feedback between all individuals will help guide future development efforts and improve and promote applications (including the newest features). To maximize its usefulness to a wide spectrum of attendants, the workshop will be split into three parts, following the three days of the event, including a (i) GAP and (M)ACE school with a basic introduction to data-driven atomistic simulation, atomic descriptors, the specifics of GAP, ACE, and MACE, and using the software packages QUIP, ACE1pack, ACEsuit/mace; a (ii) "What's new?" section with new methods and techniques for better software, scalability, parallelization, GPU porting, wrappers for automatic workflows for potential generation; a (iii) GAP/(M)ACE enabled science section with examples from various scientific fields to highlight the diverse ways of pushing the scales using GAP and (M)ACE. A main component of the workshop will be lectures which will be further complemented by contributed talks, hands-on sessions, and a poster presentation that allow more personal exchange on scientific and technical issues. To specifically attract a large number of students and early career researchers, we aim to support their participation by providing a limited number of free on-campus housing.