Effect of Increasing the Descriptor Set on Machine Learning Prediction of Small Molecule-Based Organic Solar Cells

Abstract

In this work, we analyzed a data set formed by 566 donor/acceptor pairs, which are part of organic solar cells recently reported. We explored the effect of different descriptors in machine learning (ML) models to predict the power conversion efficiency (PCE) of these cells. The investigated descriptors are classified into two main categories: structural (topology properties) and physical descriptors (energy levels, molecular size, light absorption, and mixing properties). In line with previous observations, ML predictions are more accurate when using both structural and physical descriptors, as opposed to only using one of them. We observed that ML predictions are also improved by using larger and more varied data sets. Importantly, the structural descriptors are the ones contributing the most to the ML models. Some physical properties are highly correlated with PCE, although they do not improve notably the ML prediction accuracy as they carry information already encoded in the structural descriptors. Given that various descriptors have significantly different computational costs, the analysis presented here can be used as a guide to construct ML models that maximize predictive power and minimize computational costs for screening large sets of candidates.

Data-driven approach towards identifying dyesensitizer molecules for higher power conversion efficiency in solar cells

“…31 The hyperparameter tuning of different K values is based on selecting an optimum value that results in minimum RMSE value and high accuracy. 32 In the splitting process, we used the regular spilling trend, such as 80% of molecule data for training the model and 20% of the molecule data for the testing of the system. Finally, we have chosen the molecules having PCE as low as 0.05% to as high as 12.5% to test the special test data case model.…”

Machine learning-assisted design of material properties

“…77 It is usually not obvious which of these ML techniques is best for a specific problem, so it can be advantageous to implement several different ML methods and choose the one with the best prediction accuracy. 59,71,78,79 Finally, the trained ML model that links the low-dimensional descriptors to the property of interest can be integrated into an iterative scheme to design materials with optimal properties. As illustrated in Figure 3A, this strategy has been employed to design random copolymers with targeted values of T g .…”

Machine learning for polymeric materials: an introduction

“…An accurate ML model requires inputs (features) that describe the system of interest. An accurate useful model depends on properly chosen and designed features 60,61 . Features are a wide range of items, from properties of atoms (e.g.…”