Integrating physical descriptors into crystal graph convolution neural networks for designing 2D materials with large exciton binding energies

Exciton binding energy ( E b ) is a critical parameter governing exciton stability, light–matter interaction, and electro-optical tunability in 2D semiconductors. However, accurately obtaining E b typically involves high-cost methods, including experimental techniques such as optical absorption spectroscopy and theoretical calculations using the computationally expensive GW–Bethe–Salpeter equation. Here, we develop a descriptor-embedded crystal graph convolutional neural network for efficient and reliable prediction of E b in 2D materials. Crystal structures are represented as graph-based embeddings, while ten physically motivated descriptors strongly correlated with E b are integrated to enhance model robustness in the small-data regime. Trained on 366 2D materials with BSE-calculated E b values, the model achieves a mean absolute error of 0.10 eV and an R 2 of 0.93 on the test set. Guided by the trained model, we screened 21684 2D materials and evaluated 3286 candidates after deduplication and stability filtering. Subsequent GW–Bethe–Salpeter validation of the top-ranked candidates identified 10 2D semiconductors with large E b values, indicative of strong excitonic effects. Representative materials such as GeFN and AlCdInSe 4 further demonstrate distinct relationships between structure and exciton property. This work establishes an efficient machine-learning framework for high-throughput evaluation of excitonic properties and provides physical design principles for 2D excitonic optoelectronic applications.