A Fixed-Point Neural Operator for Size- and Functional-Transferable Hamiltonian Prediction

The article presents HamEvo, a fixed-point neural operator designed for predicting the Kohn-Sham Hamiltonian, which enhances density functional theory (DFT) predictions while maintaining access to critical electronic structure observables. HamEvo achieves a 35-49% reduction in Hamiltonian errors compared to traditional methods, with mean absolute errors for HOMO and LUMO energies at 0.036 and 0.053 eV, respectively, and demonstrates effective few-shot fine-tuning for larger molecules (up to 122 atoms). Its inference speed is significantly improved, being up to 242 times faster than conventional DFT, making it a valuable tool for practitioners looking to optimize computational efficiency in electronic structure calculations.