Machine-Learning-Driven Advances in Modeling Inorganic Materials

Understanding the connections between atomic-scale structures and macroscopic properties is among the most important research challenges in solid-state and materials chemistry. Computer simulations based on the laws of quantum mechanics have played a key role in this – but they are computationally demanding, and therefore they will inevitably reach their limits when materials with highly complex structures are to be studied. Machine learning (ML) based interatomic potentials are a rapidly emerging approach that helps to overcome this limitation: being “trained” on a suitably chosen set of quantum-mechanical data, they achieve comparable accuracy whilst giving access to much larger-scale simulations – with thousands or even millions of atoms.<br/><br/>In this presentation, I will showcase some recent advances in the modeling and understanding of inorganic materials that have been enabled by ML-driven molecular dynamics (MD) simulations. I will argue that ML potentials are particularly useful for modeling structurally complex inorganic solids, such as non-crystalline (amorphous) phases that are difficult to characterize experimentally. I will survey recent work that ranges from structural studies of amorphous elements at ambient and high pressure to the modeling of materials for practical applications – for example, of disordered carbon phases for lithium- and sodium-ion battery anodes, or phase-change memories for digital data storage. I will discuss perspectives for combining these increasingly popular ML-driven simulation tools with first-principles electronic-structure computations and chemical-bonding studies, as well as possible new synergies with experimental approaches – with a vision to more fully understand, and ultimately to design, complex inorganic materials on the atomic scale.