Combining Device Information and Molecular Fingerprints into a Global Model for Reliable Prediction of Power Conversion Efficiency of Dye-Sensitized Solar Cells.

Solar cells hold the potential to meet the growing worldwide demand for clean and sustainable energy. Today most solar cells are based on silicon, yet new alternatives are continuously emerging. In particular, dye-sensitized solar cells (DSSCs) constitute low-cost electrochemical energy-harvesting devices. The design of new DSSCs with improved properties could be aided by computational techniques, in particular machine learning (ML) models. In this work, we therefore present the first global ML model based on nearly all entries (4351 out of 4426) in the DSSC Database (DSSCDB). Recognizing that the performance of DSSCs depends on both dye and device, cells were characterized by a combination of Morgan fingerprints and device features, and the resulting dye-device pairs were subjected to Histogram-based Gradient Boosting (HGB) regression using Power Conversion Efficiency (PCE) as the modeled end point. Considering both dye and device properties allowed us to resolve duplicate dye entries found in the DSSCDB and to highlight the importance of both components to cell performance. The resulting global model boasts high predictive ability on the test set (Q2 = 0.73 ± 0.02, MAE = 0.97 ± 0.03 standard deviation over 100 random splits) and has a broad applicability domain, covering organic compounds of various scaffolds, metal-organic compounds, and dye mixes. SHAP analysis was performed to elucidate structure-property relationships. The implications of this study contribute methodological and chemical novelty in the search for technology that can meet the world's energy demand.