GATFormer: A Graph Attention Transformer for Accurate Bandgap Prediction in Semiconductor Nanostructures Toward Optoelectronic Device Design

The electronic bandgap of semiconductor nanostructures fundamentally determines their suitability for optoelectronic applications including photovoltaics, light-emitting diodes, and photodetectors. However, traditional density functional theory (DFT) calculations, while accurate, remain computationally prohibitive for high-throughput screening of nanostructured materials required for device optimization. Here, we propose GATFormer, a novel Graph Attention Transformer architecture that synergistically combines graph attention mechanisms with Transformer self-attention to achieve state-of-the-art bandgap prediction for semiconductor nanostructures relevant to nanoelectronics. Our architecture introduces three key innovations tailored to nanoscale systems: (1) a hybrid graph-Transformer encoder that captures both local atomic bonding environments and long-range structural correlations critical for quantum-confined structures; (2) a nano-aware positional encoding scheme that explicitly distinguishes surface atoms from bulk-interior atoms and encodes quantum confinement directionality across different dimensionalities (quantum wells, wires, and dots); and (3) a hierarchical pre-training strategy leveraging bulk crystal data before fine-tuning on nanostructure subsets. Extensive validation on the JARVIS-DFT dataset comprising approximately 75,000 materials demonstrates that GATFormer achieves a mean absolute error (MAE) of 0.072 eV for bandgap prediction, representing a 26.5% improvement over the current state-of-the-art ALIGNN model. Notably, for nanostructured materials that dominate next-generation optoelectronic device platforms, the improvement is even more pronounced, with MAE reduced from 0.128 eV to 0.094 eV. Our approach establishes a computationally efficient framework for structure-property prediction in nanoscale semiconductors, enabling rapid virtual screening of candidate nanostructures for targeted optoelectronic applications.