Empowering Materials Science with Machine Learning: From Electron Density to Atomistic Simulations

Abstract

The development of modern society requires new materials to handle resource and environmental crises. Traditionally, the search for new materials is based on large numbers of trials and errors. Meanwhile, plenty of data, especially from failures, are wasted rather than forms experiences that can be shared among researchers. Machine learning techniques can help deal with complex information and take part in the design and exploration loop of materials, accelerating the simulation and property prediction process. In this thesis, we mainly aim to implement machine learning techniques for microscopic understanding of materials.

In Chapter 3, a study on rare events in iodide-related point defect hopping in CsPbI3 was conducted. We found that the existence of iodide vacancies or interstitials can effectively reduce the creation barrier for anti-Frenkel disorders. These additional defects can assist the fast ion diffusion, as well as the defect annihilation.

In Chapter 4, we have found and systematically investigated the planar confinement behaviour of iodide interstitials in CsPbI3. According to the molecular dynamics and kinetic Monte Carlo simulations, this behaviour will influence the defect lifetime and diffusion of ions. Moreover, we have also studied the influence of symmetry breaking in low-temperature phases. The interstitials obtain higher on-site energy in the equatorial planes. This can further restrict the motion of interstitials, preventing them hop inside the equatorial plane.

In Chapter 5, the Deep Charge model was developed and described. The model can predict electron density from the input structures with ab initio accuracy. The model has been tested in various materials systems. The machine learning techniques provide a way to represent electron density with a linear time complexity, rather than the cubic complexity of density functional theory calculations. This can facilitate fast calculation on larger systems and the physical property predictions.