Data-driven designing of dyes: Chemical space generation and dipole moment prediction

Abstract

The current study presents machine learning-assisted designing of dyes for photovoltaics applications. Multiple machine learning models are trained to predict the dipole moment. Random forest model has appeared as best model with lower root mean square error value (1.01 Debye) and higher r-squared value (0.87). New dyes are designed using automatic method and their dipole moment is predicted using best machine learning model. The generated chemical space of dyes is visualized and analyzed using cluster plot, silhouette plot and t-distributed Stochastic Neighbor Embedding (t-SNE plot). 30 dyes with highest dipole moment values (6.31–7.12 Debye) are chosen. Chemical similarity analyses are performed on the selected dyes using cluster analysis and heatmap. Furthermore, an investigation into the synthetic accessibility score of the newly designed dyes is conducted. This method facilitates the swift selection of dyes for potential use in photovoltaic devices.