Machine learning-based q-RASPR modeling of power conversion efficiency of organic dyes in dye-sensitized solar cells

Abstract

Different computational tools are now popularly used as an alternative to experiments for predicting several property endpoints of industrial importance. Recently, read-across and quantitative structure–property relationship (QSPR) have been merged to develop a new modeling technique read-across structure–property relationship (RASPR) which appears to have much potential in predictive modeling. This approach is also promising for modeling relatively smaller data sets as the similarity-based RASPR descriptors are computed from multiple structural and physicochemical features. To understand the potential of RASPR in data gap filling, we have undertaken a case study of modeling Power Conversion Efficiency (PCE) of different classes of organic dyes used in Dye-Sensitized Solar Cells (DSSCs) for renewable energy generation. We have used a large dataset of 429 compounds covering 4 classes of organic dyes. We initially performed read-across analysis using different similarity measures with structural analogues for query compounds and calculated the weighted average predictions. Based on the read-across optimized settings, RASPR descriptors were calculated, and these were then merged with the chemical descriptors, and finally, a single partial least squares (PLS) model was developed for each of the dye classes after feature selection, followed by additional Machine Learning (ML) models. The external prediction quality of the final RASPR models superseded those of the previously developed QSPR models using the same level of chemical information. The important structural features and similarity measures contributing to the PCE have been extracted using the RASPR method which can be used to enhance the PCE values in the newly designed dyes. The RASPR method may also be efficiently applied in modeling other properties of interest in a similar manner.