Generating a vast chemical space for high polar surface area triphenylamine polymers by machine learning-DFT calculations assisted reverse engineering for photovoltaics

Abstract

The total polar surface area (TPSA) is a crucial parameter in photovoltaic (PV) materials, as it directly influences their solubility, processability, and device performance. This study leverages machine learning-assisted reverse engineering to generate a vast chemical space of high polar surface area triphenylamine (TPA) polymers for PV applications. By applying co-gradient boosted (xGBoost) and Random Forest algorithms to a dataset of 543 triphenylamine-based chromophores, high accuracy (R2 = 0.93–0.96) is achieved in predicting the TPSA of these chromophores. Feature importance analysis using the Shapley Additive eXplanation (SHAP) values reveals that the number of nitro groups (NOCount) has the highest impact on model performance. The generated model is rigorously evaluated using K-fold cross-validation and out-of-bag evaluation. 1000 new polymers are then generated with predicted TPSA values, including some with exceptionally high TPSA of up to 182. Further analysis of the charge transfer patterns in selected polymers shows that the highest occupied molecular orbital (HOMO) and lowest unoccupied molecular orbital (LUMO) are oriented in opposite directions, indicating a high potential for these materials in PV devices. The predicted PV performance of these polymers exhibits promising characteristics, with values of 54–79 % for the light-harvesting efficiency (LHIE), 1.63–1.68 V for the open-circuit voltage (V_oc), 0.54–0.92 for the fill factor (FF), and 21.99–32.43 mA/cm² for the short-circuit current density (J_sc).