Deep-learning-assisted photovoltaic performance prediction of sensitizers in dye-sensitized solar cells

Abstract

Research on efficient sensitizers for dye-sensitized solar cells (DSSCs) is in high demand to further improve the device efficiency, which can be accelerated through computational protocols. However, characterizing the photovoltaic properties of a sensitizer relies on expensive theoretical simulations for molecular excited states, making it a great challenge to achieve cost-effective performance prediction. In this study, an accurate and robust model is established by combining deep learning (DL) technique and quantum chemical (QC) calculation. By exclusively utilizing ground-state-related features as initial descriptors, a feedforward neural network is employed to predict the averaged light-harvesting efficiency and photocurrent density associated with absorption for dyes in DSSCs. The DL algorithm achieves a Pearson correlation coefficient exceeding 0.9 when compared to QC results, qualitatively reproducing experimental observations and effectively revealing relevant molecular design principles. Furthermore, the reliability of this approach is validated by the assessment of six newly designed molecules. The DL prediction exhibits a high consistency with the calculated result, indicating its potential in the discovery of prospective materials. Remarkably, the DL method shows a substantial advantage in terms of speed, being capable of calculations that are millions of times faster than comparable QC calculations, which provides an efficient tool for identifying candidates with preferable photovoltaic performance. The proposed model is expected to contribute to the rapid screening of sensitizers with desired photo-physical properties and to discovering new design strategies for high-performance DSSCs.