Developing machine learning models that generalize across diverse chemical compositions remains a fundamental challenge in electronic structure theory. Conventional data-driven approaches often exhibit limited transferability due to their reliance on memorizing potential energy surfaces of specific materials. In this work, we present a Slater-Koster (SK) constrained graph attention neural network, implemented as the Graph Attention Tight-Binding (GATB) model, to learn transferable orbital interaction rules. Unlike standard geometric message passing, our framework explicitly embeds the classical SK tight-binding formalism into the neural architecture as a hard constraint. Specifically, we employ orthogonal Bessel basis functions to parametrize the radial decay of SK bond integrals, enforcing rotational invariance and physical asymptotic behavior. This physics-embedded design forces the model to infer electronic couplings based on atomic species and spatial arrangements rather than overfitting to specific geometries. We demonstrate that GATB achieves robust, zero-shot generalization in constructing Hamiltonians and predicting band structures for transition metal dichalcogenides (TMDCs) with chemical compositions unseen during training. Finally, we deploy the trained model as an open-access online tool to facilitate real-time cloud-based Hamiltonian and electronic structure exploration.

Close

PrevNext

View figure in article

• Electronic structure
• First-principles calculations
• Machine learning
• Neural network simulations
• Tight-binding model