Bridging predictive machine learning modeling with experimental fabrication of organic solar cells

Abstract

Integrating Machine Learning into organic photovoltaic research offers a systematic framework toward efficient parameter screening; however, the physical validation of these computational models during manufacturing remains a significant obstacle. This study proposes a “Predict-then-Verify” workflow designed to bridge the gap between materials informatics and sustainable laboratory-scale fabrication. While previous models demonstrated a reliable ability to predict unseen devices from external datasets, their capacity to screen optimal compositions had not yet been validated in a real manufacturing environment. We take advantage of these Machine Learning models to design and physically validate P3HT:PCBM solar cells. We also achieved the optimal device configuration in just two experimental iterations, reducing material consumption from the traditional trial and error process. To isolate intrinsic degradation mechanisms from atmospheric variables, encapsulated devices were evaluated over a 30-day temporal horizon. Power conversion efficiency predictions from Gradient Boosting, Random Forest, Neural Networks and Multivariate Linear models were evaluated against the fabricated cells. The Gradient Boosting approach achieved optimal short- to medium-term predictions with relative errors around 5%, while late-stage degradation boundaries (below 80% initial efficiency) remained consistently bounded within a 20–30% divergence margin. To ensure full reproducibility, all datasets and documentation are openly accessible on GitHub.