GAN-augmented machine learning enables accurate band gap prediction for 2D lead iodide perovskites with limited data.

Low-dimensional hybrid lead iodide perovskites exhibit band gaps that are highly sensitive to subtle octahedral distortions, yet accurate prediction remains challenging under small-data regimes where traditional machine learning models tend to fail. Herein, we develop a collaborative machine-learning framework for two-dimensional (2D) lead iodide perovskites that integrates physically interpretable [PbI6]4--based structural descriptors, principal component analysis (PCA) for dimensionality reduction, multi-layer perceptron generative adversarial network (MLP-GAN) data augmentation (generating 1000 synthetic structures), and automated hyperparameter optimization. Using 107 single-crystal experimental data points, we benchmark nine regression models and demonstrate that GAN-based augmentation substantially improves model learning capability and generalization robustness. This study is designed to establish an interpretable data-augmentation strategy for small-data materials modeling and to test its applicability to band-gap prediction in 2D lead iodide perovskites.