Physics-aware graph neural networks for automated tight-binding model construction in quantum transport simulations

Tight-binding (TB) model is crucial for quantum transport simulations of semiconductor devices, critically determining the electrical characteristics of channel materials. Here, we propose a graph neural network (GNN)-based framework for automated TB model construction. By integrating atomic sites (nodes) and chemical bonds (edges) into atomistic graph representations, our method efficiently extracts orbital onsite energies and inter-orbital hopping parameters through supervised learning of density functional theory-derived band structures. The physics-aware architecture of GNNs, which inherently mirrors the atomic and bonding configurations of materials, ensures that the predicted TB models retain intrinsic physical interpretability, including the sparse matrix form, tunable orbital localization, exponentially decaying hopping strength, and defect-resolved local density of states, significantly broadening their applicability. As a result, it allows co-training on defective and defect-free systems, so that structural perturbations are naturally encoded as parameter changes, overcoming the lack of hopping parameters between distinct configurations and enabling the construction of Hamiltonians for non-periodic, defect-containing device channels, a longstanding challenge in ab initio quantum transport modeling. Furthermore, we develop the model size scaling and band structure editing functionalities, enabling flexible manipulation of electronic properties and cutting computational costs. We apply this framework to amorphous In-Ga-Zn-O and 4H-silicon carbide, two technologically critical channel materials whose performance-limiting defects cause unresolved reliability issues such as current degradation and threshold voltage drift, arising from quantum effects that cannot be captured by conventional drift-diffusion models. This work bridges the gap between ab initio calculations and device-level modeling, offering a transformative tool for semiconductor device design and beyond.