Machine learning analysis on performance of naturally-sensitized solar cells

Abstract

Understanding natural photosensitizers characteristics in dye-sensitized solar cells (DSSCs) is necessary to achieve high power conversion efficiency (PCE). Here, we statistically investigate the possibility to achieve relatively high PCEs in naturally-sensitized-photoanode-based DSSCs using decision trees and support vector training of dye structural, electronic, and molecular properties. We studied the chemical structure and bandgap of 27 sensitizers, correlated to the literature reported PCEs while applying “in-between randomization” for datasets expansion. Training and testing algorithms were carried out via 4 (dye) predictors including the number of π-bonds (PI), anchoring groups (X), HOMO(H)-LUMO(L), and bandgap energy (BG), with 2 responses for the possibility to achieve PCEs >1.82% (Yes/No). Both HLBG-input and PIXBG-input models were found promising with the highest accuracies of ∼ 92% using trees classifiers and ∼ 96% with support vector classification, respectively. Testing datasets were chosen to check for built models accuracy and similarities were evident between the models' results (PIX, BG, HLBG, and PIXBG). Residual analysis showed trees models had the minimum statistical errors with narrow violons (± 0.25 ranges). Despite that the existence of more anchoring groups allows firm molecules attachment to semiconductors for enhanced charge injection, the HLBG-input analysis confirmed that BG is the foremost controlling parameter (∼ 3-fold > H), with BG/X importance ratio of 12 leading to the parameter's importance: BG (1) > H (0.32) > PI (0.08) > X (0.04). This work shows the potential of adopting trained classifiers for analyzing natural sensitizer's abilities to inject and separate generated electron-hole pairs for producing renewable, cost-effective, and sustainable energy.