Beyond Zero Kelvin: Machine-Learned Potentials for Finite-Temperature Modeling of Semiconductors

Licentiate thesis, 2026

The global transition toward sustainable energy sources hinges on the development of efficient, stable and cost-effective photovoltaic materials.
Halide perovskites have emerged as leading candidates, with power conversion efficiencies exceeding 25%, yet their operational stability remains a critical bottleneck. Addressing this challenge requires theoretical modeling, but standard density-functional theory (DFT) is largely restricted to zero-kelvin simulations, and extending first-principles methods to finite temperatures remains computationally prohibitive. Machine-learned potentials (MLPs) dramatically reduce this computational cost, enabling finite-temperature simulations at DFT-level accuracy.

This thesis applies this framework to two classes of problems. First, neuroevolution potential (NEP) models are used to decipher the phase diagrams and structural dynamics of the mixed perovskite MA_1-xFA_xPbI₃, including the identification of a morphotropic phase boundary, and to investigate the debated low-temperature $\gamma$-phase of formamidinium lead iodide (FAPbI₃). Second, the same framework is extended to the thermodynamics of charged point defects, where thermodynamic integration (TI) is combined with NEP models to evaluate defect formation free energies and charge transition levels as a function of temperature, revealing that thermal effects can substantially shift these quantities across a range of semiconductors.

Taken together, these contributions advance the development of predictive, temperature-aware models of technologically relevant materials, representing a necessary step toward the rational design of stable and efficient next-generation solar cells.