This repository contains the source code and manuscript for our study on predicting the long-term photoluminescence (PL) stability of metal halide perovskites.
We present a hybrid CNN-LSTM time-series forecasting algorithm that predicts long-term stability trends using short-term input data derived from alloyed nanocrystals. By employing physics-informed featurization, this approach outperforms naïve benchmarks and ensures intrinsic interpretability through SHAP analysis.
Manuscript_Physics-informed time-series forecasting of perovskite photoluminescence stability.pdf: The manuscript detailing the methodology, results, and discussion.main_featurization_pipeline.ipynb: The automated spectral segmentation algorithm that extracts physically meaningful, time-resolved features from raw PL spectra (e.g., HE/LE peak separation)main.ipynb: The primary training script for the hybrid CNN-LSTM architecture.TrainTestSplit.ipynb: Utilities for splitting datasets by composition (8:2 ratio) and applying the windowing approach to ensure time-invariance.model_testing.ipynb: Evaluation scripts for calculating RMSE against naïve benchmarks.