Accelerated Discovery of Efficient Solar Cell Materials Using Quantum and Machine-Learning Methods

Abstract

Solar-energy plays an important role in solving serious environmental problems and meeting highenergy demand. However, the lack of suitable materials hinders further progress of this technology. Here, we present the largest inorganic solar-cell material search to date using density functional theory (DFT) and machine-learning approaches. We calculated the spectroscopic limited maximum efficiency (SLME) using Tran-Blaha modified Becke-Johnson potential for 5097 non-metallic materials and identified 1997 candidates with an SLME higher than 10%, including 934 candidates with suitable convex-hull stability and effective carrier mass. Screening for 2D-layered cases, we found 58 potential materials and performed G0W0 calculations on a subset to estimate the prediction-uncertainty. As the above DFT methods are still computationally expensive, we developed a high accuracy machine learning model to pre-screen efficient materials and applied it to over a million materials. Our results provide a general framework and universal strategy for the design of high-efficiency solar cell materials. The data and tools are publicly 2 distributed at:

A Critical Review of Machine Learning of Energy Materials

“…Using 256 GW calculations, the authors were able to identify in this group high‐SLME materials including almost all contemporary PV absorber materials. Choudhary et al have performed meta‐GGA calculations of bandgaps using the Tran–Blaha‐modified Becke–Johnson (TBmBJ) potential for 12 881 out of ≈30 000 materials in the JAVIS‐DFT database . The SLME metric was calculated on 5097 nonmetallic materials and the elemental distribution for high‐SLME materials is shown in Figure 8 a.…”

“…The SLMEs with direct bandgap values are shown in Figure 8b, where a consistent volcano shape as original SLME results using GW is obtained. A threshold of 10% was chosen to label high‐SMLE materials versus low‐SMLE materials, and the binarized data formed the training data set for ML modeling . The authors then used CFID as structure features and compared various classification models in terms of the classification AUC.…”

“…a) Element distribution for high‐SLME materials, b) the SLME distribution with direct bandgap values, and c) ROC curves for a GBDT model. Adapted with permission . Copyright 2019 American Chemical Society.…”

“…[7,66,115,116] The goal of such ML models is typically to bypass expensive and time-consuming experiments or firstprinciples computations. They have been used to provide guidance to experimental design, i.e., the next areas to explore for a potential "blockbuster" material for an application [117][118][119][120][121] as well as rapid computational screening of chemical spaces. [30,[122][123][124][125] ML-IAPs are also a subcategory of supervised materials ML models where the target is to predict the energies, forces and stresses, i.e., the potential energy surface, for a given atomic configuration.…”

“…A threshold of 10% was chosen to label high-SMLE materials versus low-SMLE materials, and the binarized data formed the training data set for ML modeling. [120] The authors then used CFID as structure features and compared various classification models in terms of the classification AUC. The AUC for decision tree models, RF, kNN, MLP, and gradient boosting decision trees (GBDT) are 0.67, 0.79, 0.77, 0.80, and 0.87, respectively.…”