Machine-Learned Interatomic Potentials for Localized Charges, Mixing, and Degradation in Energy Semiconductors
Abstract
Modern energy semiconductors such as halide perovskites and oxide photoanodes are soft, chemically mixed, and rich in defects. Their performance in solar cells, light emitters, and photoelectrochemical devices is therefore determined by processes that occur at operating temperature on length and time scales far beyond direct density-functional molecular dynamics: the motion of localized charges (polarons and self-trapped excitons), phase transformations driven by chemical mixing, and degradation at surfaces and interfaces.
The goal of this project is to resolve these processes with machine-learned interatomic potentials of the neuroevolution-potential (NEP) family. The models are neural networks trained on density-functional reference data through active learning, using NEP, the charge-aware qNEP developed in the group, and tensorial NEPs for dielectric and Raman response, all implemented in the open-source GPUMD package. With these models we can simulate 10000-1000000 atoms over nanoseconds with close to DFT accuracy. The group has used them to identify the low-temperature phase of FAPbI3, map a morphotropic phase boundary in mixed-cation lead iodide perovskites, and follow the coupled dynamics of oxygen vacancies and polarons in BiVO4. In this project we extend the models to charged and excited states, to interfaces between solids, liquids, and molecules, and to response properties, and apply them to three problems: how polarons and self-trapped excitons behave at finite temperature and at the BiVO4/water interface; how mixing in Pb-, Sn-, and Ge-based halide perovskites controls phase behaviour, including the transformation to the non-perovskite phase that limits device stability; and how perovskite films degrade in contact with water and organic molecules, and how adsorbed molecules passivate or protect them. Both training and molecular dynamics run entirely on GPUs.
The results are compared directly with NMR, Raman, infrared, and stability measurements from our experimental collaborators. Meeting the goal will give atomistic criteria for stabilizing perovskite films and interfaces against degradation, for choosing compositions that stay in the photoactive phase, and for controlling charge trapping in photoanodes, which are the main obstacles to durable perovskite solar cells and efficient solar-fuel devices.
The project is part of the PI's Wallenberg Academy Fellowship and KAW project "From Atoms to Devices: Understanding Interface Phenomena for Better Perovskite Optoelectronics", and complements the group's NAISS Large project on Arrhenius, which provides the electronic-structure reference data.
