Inverse Structure Determination with Atomistic Machine Learning Tools

For amorphous materials, the primary modeling problem is the elucidation of the atomic-scale structure, from which all material properties follow. Because for amorphous materials the structure is not well defined and does not generally correspond to thermodynamic equilibrium, the sampling problem becomes rather convoluted. Direct simulation, without experimental input, can lead to incorrect structures. Adding experimental constraints to the structure determination, on the other hand, offers a route to enforce experimental agreement by design, as is done in such techniques as reverse Monte Carlo (RMC). However, there is a vast ensemble of structures that agree with, say, the diffraction pattern of a material, some of which are unrealistically high in energy. To further constrain the search space, energetic constraints, based on an explicit description of the potential energy surface, can be added, as in hybrid RMC (HRMC) [1]. With the advent of atomistic machine learning technology, these techniques can be extended to better potential energy models and experimental observables beyond diffraction [2]. Furthermore, these techniques accept computationally efficient analytical gradients of the generalized objective function that leads to experimental agreement, enabling linear-scaling methods that far exceed the performance of quadratic MC sampling [3,4]. In this presentation, I will try to convince you of the usefulness of these techniques for explaining the structure of amorphous materials.

[1] G. Opletal, T. Petersen, B. O'Malley, I. Snook, D.G. McCulloch, N.A. Marks, and I. Yarovsky. Mol. Simul. 28, 927 (2002).
[2] T. Zarrouk, R. Ibragimova, A.P. Bartók, and M.A. Caro. J. Am. Chem. Soc. 146, 14645 (2024).
[3] T. Zarrouk and M.A. Caro. arXiv:2508.17132.
[4] T. Zarrouk and M.A. Caro. arXiv:2509.22388.