GPU-Accelerated Machine Learning for Analysis of Time-resolved Photoluminescence Data

Quantifying charge-carrier dynamics within a material or device from analysis of optoelectronic measurements is a crucial aspect of improving next-generation solar cells. However, analyses by hand are limited in information yield due to their reliance on simplified physics models. Simulation of full physics models can be computationally expensive. Here we demonstrate a GPU-accelerated machine learning approach via Bayesian parameter estimation for rapid analysis of optoelectronic data. Using time-resolved photoluminescence (TRPL) data of a perovskite absorber as a case study, we demonstrate our ability to estimate carrier mobilities, the doping level, and the radiative recombination rate. Furthermore, while most TRPL analyses are limited to determining an effective minority carrier lifetime, we reliably decompose this recombination lifetime into radiative, bulk nonradiative, and surface nonradiative components by introduction of TRPL data from multiple absorbers of different thicknesses. Our simultaneous collection of these parameters represents a significant increase in the typical information yield from TRPL measurements.