Abstract

Machine learned interatomic potentials (MLIPs) have become essential tools for modeling condensed matter systems with near ab initio accuracy across extended length and time scales. This mini-colloquium will highlight recent developments that merge physical insight with advanced strategies for machine-learning inter-atomic interactions. Symmetry equivariant neural networks, transformer architectures, moment-tensor potentials, Gaussian-approximation potential-type approaches, and physically constrained graph based frameworks have expanded the applicability of MLIPs to complex crystalline, amorphous, and low dimensional materials. Increasingly, progress is defined less by raw accuracy, which often already surpasses the chemical accuracy threshold, than by advances in computational efficiency, stability, and generalizability. Additional topics include coupling ML force fields to electronic excited state dynamics and extending accessible time and length scales through coarse graining and adaptive resolution approaches. Discussion will address strategies for advanced MLIP parametrization, the choice of suitable benchmarks, uncertainty quantification, and the development of transferable, interpretable models for complex materials. A central topic of the symposium will also be, how MLIPs in the past years have lead to a true paradigm shift in the simulation of various physically relevant properties of both traditional as well as advanced materials systems. By bringing together experts from theory, machine learning, and computational materials physics, the colloquium aims to define the next directions in physically grounded and computationally efficient atomistic modeling.

Organizers