Machine-learning potentials for finitetemperature simulations of polarons

Abstract

The performance of functional materials in advanced energy technologies, such as photovoltaics and photocatalysis, can be strongly influenced by charge localization. Excess charges can cause distortions in the surrounding lattice and localize in a self-generated potential well, forming quasiparticles called polarons. Although firstprinciples methods based on density functional theory (DFT) offer accurate results for polaron calculations, these methods are limited at finite temperatures due to their high computational cost. In the present work, polarons are modelled in MgO and LiF using machine-learned interatomic potentials (MLIPs) to enable finite-temperature calculations at high speed. At 0 K, LiF exhibits two types of stable polarons. The transition between the two polaron states is first studied with hybrid density functional calculations using the string method. One of the polaron states is metastable and the energy barrier is found to be small, 27 meV, making it difficult to model the metastable polaron state with MLIPs. The more stable polaron state is then studied at finite temperatures, together with the polaron in MgO. Using hybrid density functional calculations as reference data, neuroevolution potential models are trained for each material. The models are used to investigate the temperature dependence of the charge transition level (CTL), a key energetic property of polarons. Temperatures up to 500K are studied, and the models predict variations in the CTL of approximately 55 meV for MgO, and 380 meV for LiF.