Machine learning for perovskite solar cells improves with provenance grouping

Provenance-grouped protocol for evaluating machine learning in perovskite photovoltaics

Machine Learning

Summary

Many machine-learning models used to predict how well perovskite solar cells work rely on random data splits that can trick the models into remembering old data instead of truly learning new physics. The authors found that many devices share identical recipes, causing this problem. They created a new way to split the data by grouping it according to where it came from, like the lab or institution, so the models must learn to predict genuinely new cases. Their tests showed that some model types only do well when they see familiar data, while others work better on truly new experiments. This approach helps separate memorization from real scientific discovery in solar cell research.

What this means in practice

  • For solar panel developers: Evaluate machine learning models on perovskite solar cell data to find designs that generalize well beyond existing recipes and institutions.
  • For materials engineers: Use provenance-grouped evaluation to better understand how fabrication methods impact solar cell efficiency predictions.

Authors

Qiwei Han, Chi Zhou

Abstract

Machine-learning models for perovskite solar cells are typically evaluated using random data splits, which we show systematically leak information through massive internal duplication. Because nearly two-thirds of curated devices share their exact recipe with another, random splits primarily reward the mere recall of historical data rather than testing true physical generalisation. To solve this, we mapped the institutional provenance of the Perovskite Database and established a rigorous hierarchy of grouped splits that progressively isolate duplicate devices, publications, and institutions. Benchmarking thirteen models establishes new state-of-the-art performance across all splits and reveals a critical regime shift: while tree ensembles dominate random splits by recognising familiar recipes, their ranking collapses on genuinely unseen institutions and future discoveries, where composition-aware neural architectures excel. Furthermore, feature importance demonstrates that the device architecture and fabrication methodology dictate final efficiency just as strongly as the absorber chemistry. This standardised protocol empowers the community to separate memorisation from true physical discovery.
View PDFOpen arXiv