Combining Automated, In Situ Experiments with Machine Learning to Quantify the Dynamic Behavior of Halide Perovskites

The enthusiasm of the scientific community towards identifying alternative materials for photovoltaics has led to the quick development of halide perovskite solar cells. Yet, despite the promising properties of these materials, they lack stability upon exposure to outdoor, environmental conditions. Because this burgeoning class of material entails a colossal chemical composition space, machine learning (ML) is very suitable to replace the conventional trial-and-error approach used in their characterization. In this talk, I will discuss how we have implemented ML models varying from echo state networks to statistical models to classify and predict physical properties such as hole transport layer electrical conductivity, halide perovskite photoluminescence response, the power conversion efficiency of photovoltaic devices, etc. Through automated, in situ optical measurements, we were able to predict the response of these materials for >50 hours, with >90% accuracy. Our high-throughput measurements and ML-supported analyses validate the potential of ML to detect and forecast hybrid perovskites’ response with a variety of chemical compositions.