From Machine-Learning Interatomic Potentials to Materials Chemistry
Understanding the links between atomic structure, chemical bonding, and macroscopic properties in materials is a formidable task. Quantum-mechanical atomistic simulations, prominently based on density-functional theory (DFT), have played important roles in this – but they are computationally expensive, and can describe complex materials only in small model systems.
Novel interatomic potentials based on machine learning (ML) have recently garnered a lot of attention in the computational materials-science community: they achieve close-to DFT accuracy but at only a fraction of the cost. In this talk, I will argue that ML-based interatomic potentials are particularly useful for studying materials with complex structures, such as amorphous (non-crystalline) solids.
I will first describe an ML potential for amorphous carbon [1] using the Gaussian Approximation Potential (GAP) framework [2], with a special view on what is needed to validate ML potentials for the amorphous state. I will then present an application to porous and partly "graphitized" carbon structures, which are relevant for applications in batteries and supercapacitors [3]; this includes a new ML strategy for simulating the movement of Li ions in such materials [4].
Finally, I will present recent work on amorphous silicon (a-Si), another prototypical non-crystalline material, where ML-driven simulations allowed us to unlock long simulation times and accurate atomistic structures [5], again making steps toward the routine and realistic atomic-scale modelling and understanding of the amorphous state.