Physics-informed time-series forecasting of perovskite photoluminescence stability
Accelerated ageing using elevated temperatures and illumination is one of the most common methods to rapidly study the stability of novel semiconductor materials. However, as the pace of materials discovery continues to accelerate, even faster stability evaluations are needed. A physics-informed time-series forecasting algorithm designed to predict the long-term photoluminescence stability of metal halide perovskites is presented. A diverse experimental dataset of 167 metal halide perovskites is collected, including different crystallinities and compositions. These are stressed using heat and light, while the photoluminescence (PL) is monitored. The >86k collected PL spectra are featurized using a physics-informed model, and a hybrid CNN-LSTM model is trained to forecast the PL intensity during degradation of samples unseen during model training. Notably, the approach generalizes across the material groups and outperforms baseline benchmarks. Furthermore, the physics-based featurization ensures explainability, enabling analysis to identify critical stability descriptors for given predictions. It is expected that this approach will be adapted to other types of time-series data and enables a pathway to significantly reduce experimental testing times.