Machine Learning Driven Studies of Performance Degradation in a-Si:H/c-Si Heterojunction Solar Cells

a-Si:H/c-Si heterojunction solar cells hold the efficiency world record around 27%, yet their market penetration is delayed. One concern is the migration of passivating hydrogen away from the interface, that some suspect may speed up the degradation of their performance. Mitigating the performance degradation necessitates the understanding of the structural evolution of a-Si:H/c-Si structures, with a focus on hydrogen migration. To this end, we have developed the SolDeg structural simulation platform that is capable of capturing extremely slow degradation processes. SolDeg integrates molecular dynamics methods that optimize the Si structure with femtosecond time steps, with the nudged elastic band method that captures the defect generation on time scales extending to gigaseconds. The molecular dynamics layer of SolDeg requires a high quality Si-H interatomic potential. While classical parametric interatomic potentials have been used extensively, the recent development of machine-learning driven interatomic potentials ignited the ambition of achieving DFT-level accuracy with classical molecular dynamics simulations. In this paper we report the development of the first machine-learning driven Gaussian Approximation Potential (GAP) to describe Si-H interactions. This potential will be used in the SolDeg platform to determine the performance degradation of a-Si:H/c-Si heterojunction solar cells.