Enhanced Design of Kesterite Solar Cells through High-Throughput Screening and Machine Learning Approaches

Kesterite Cu2ZnSnS4 (CZTS) is regarded
as
one of the most promising materials for thin-film solar cells due
to its high light absorption capability, composition of earth-abundant
and nontoxic elements, and ease of low-cost mass production. Although
the certified power conversion efficiency (PCE) of kesterite solar
cells has exceeded 14%, this efficiency remains significantly below
the Shockley–Queisser limit. In this study, we generated a
Perdew–Burke–Ernzerhof (PBE) band gap data set encompassing
263 64-atom species for high-throughput screening by substituting
elements at different sites in A2BCX4 quaternary
kesterite materials. Additionally, we utilized a symbolic regression
method based on genetic programming to explore the functional relationship
among the oxidation state, ionic radius, and electronegativity of
kesterites with PBE band gaps. Simultaneously, we employed decision
tree models (XGBoost, LightGBM, CatBoost, and random forest) and convolutional
neural network (CNN) models (CustomCNN, VGG16, DenseNet121, Xception,
and EfficientNetV2B0) to predict band gaps, achieving a coefficient
of determination (R2) of up to 0.93. Furthermore,
we selected 54 kesterite materials with PBE band gaps ranging from
0.4 to 1.5 eV for detailed electronic structure calculations with
Heyd–Scuseria–Ernzerhof (HSE06) functional and investigated
the effects of B-site atomic substitutions on the performance of solar
cell materials. Compared to Ag2CaSnSe4, Ag2SrSnSe4 exhibits fewer deep defects and richer
shallow defects, which contribute to an increased carrier concentration
and reduced charge and energy losses, making it a superior candidate
for solar cell applications.