Strategic Integration of Machine Learning in the Design of Excellent Hybrid Perovskite Solar Cells

The photoelectric conversion efficiency (PCE) of perovskites
remains
beneath the Shockley-Queisser limit, despite its significant potential
for solar cell applications. The present focus is on investigating
potential multicomponent perovskite candidates, particularly on the
application of machine learning to expedite band gap screening. To
efficiently identify high-performance perovskites, we utilized a data
set of 1346 hybrid organic–inorganic perovskites and employed
11 machine learning models, including decision trees, convolutional
neural networks (CNNs), and graph neural networks (GNNs). Four descriptors
were utilized for high-throughput screening: sine matrix, Ewald sum
matrix, atom-centered symmetry functions (ACSF), and many-body tensor
representation (MBTR). The results indicated that LightGBM and CatBoost
somewhat surpassed XGBoost in decision tree models, but random forests
lagged. Among the CNN models utilizing the same four descriptors,
CustomCNN and VGG16 surpassed Xception, while EfficientNetV2B0 exhibited
the least favorable performance. When the sine matrix and Ewald sum
matrix served as adjacency matrices in GNN models, GCSConv exhibited
a considerable improvement over GATConv and a slight advantage over
GCNConv. Significantly, GCSConv outperformed other models when utilized
with the Ewald sum matrix. The ideal combination of descriptors and
algorithms identified was MBTR + CustomCNN, with an R2 of 0.94. Subsequently, three perovskites exhibiting
appropriate Heyd–Scuseria–Ernzerhof (HSE06) band gaps
were identified to define the defects. Among them, CH3C(NH2)2SnI3 exhibited superior performance
in both vacancy and substitutional defects compared to C3H8NSnI3 and (CH3)2NH2SnI3. This high-throughput screening method with
machine learning establishes a robust foundation for selecting solar
materials with exceptional photoelectric properties.