Where the Error Lives: Spectral Representations of Learned Kohn-Sham Hamiltonians
Abstract
Kohn-Sham density functional theory (DFT) is the standard method for computing the electronic structure of molecules and materials, but its self-consistent-field (SCF) iterations make it expensive for large systems and high-throughput screening. Machine-learned Hamiltonian models offer a shortcut: an equivariant network predicts the Kohn-Sham Hamiltonian matrix in one forward pass, and diagonalizing the prediction yields orbital energies, the HOMO-LUMO gap, molecular orbitals and the electron density. These models are trained to match the reference matrix entry by entry, yet a small entry-wise error does not guarantee accurate eigenvalues or eigenvectors. Two physics-informed alternatives go partway: a recently proposed wavefunction-alignment loss adds a spectral penalty without diagonalizing the prediction, and matching the electron density of the diagonalized prediction supervises it directly. Neither is complete: by representing the prediction error in the eigenbasis of the reference Hamiltonian, we show that each sees only part of the error, and that shifting the predicted orbital energies leaves the electron density exactly unchanged. The same analysis identifies a complete training signal: the cumulative local density of states (cumulative LDOS), the number of electrons at each point in space with energy below a given threshold. The cumulative LDOS is a single physical observable that contains the electron density as a special case, responds to every component of the error, determines the Hamiltonian uniquely, and yields gradients that do not vanish for large orbitalenergy errors. Fine-tuning three Hamiltonian networks on QH9 in this way, which we call FermiScope, reduces the HOMO-LUMO gap error without increasing the element-wise error. The gains carry over to out-of-distribution and molecular-dynamics splits, where the wavefunction-alignment loss helps less.
Related
- Hybridizing expert descriptors and learned embeddings yields transferable Mat2Vec-* element representations
- Machine-Learned Dynamical Representations for Accelerated RiteWeight Convergence
- The Interference Index: Quantifying and Improving Error Cancellation in Machine-Learned Thermodynamic Stability Predictions
- ALIGNing Learned Representations: Domain-General Models Triangulate Physicochemical Interactions in Analytical Separations
- Spectral symmetry fingerprints in equivariant representations
BibTeX
@misc{sun2026where,
title = {{Where the Error Lives: Spectral Representations of Learned Kohn-Sham Hamiltonians}},
author = {Fang Sun and Hekai Cheng and Ruoyan Li and Yuanzhou Chen and Wei Wang and Jason Cong and Yizhou Sun},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.37843},
eprint = {2610.37843},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.37843},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Where the Error Lives: Spectral Representations of Learned Kohn-Sham Hamiltonians
AU - Fang Sun
AU - Hekai Cheng
AU - Ruoyan Li
AU - Yuanzhou Chen
AU - Wei Wang
AU - Jason Cong
AU - Yizhou Sun
PY - 2026
DA - 2026/10/27
PB - arXiv
DO - 10.48550/arxiv.2610.37843
UR - https://arxiv.org/abs/2610.37843
AB - Kohn-Sham density functional theory (DFT) is the standard method for computing the electronic structure of molecules and materials, but its self-consistent-field (SCF) iterations make it expensive for large systems and high-throughput screening. Machine-learned Hamiltonian models offer a shortcut: an equivariant network predicts the Kohn-Sham Hamiltonian matrix in one forward pass, and diagonalizing the prediction yields orbital energies, the HOMO-LUMO gap, molecular orbitals and the electron density. These models are trained to match the reference matrix entry by entry, yet a small entry-wise error does not guarantee accurate eigenvalues or eigenvectors. Two physics-informed alternatives go partway: a recently proposed wavefunction-alignment loss adds a spectral penalty without diagonalizing the prediction, and matching the electron density of the diagonalized prediction supervises it directly. Neither is complete: by representing the prediction error in the eigenbasis of the reference Hamiltonian, we show that each sees only part of the error, and that shifting the predicted orbital energies leaves the electron density exactly unchanged. The same analysis identifies a complete training signal: the cumulative local density of states (cumulative LDOS), the number of electrons at each point in space with energy below a given threshold. The cumulative LDOS is a single physical observable that contains the electron density as a special case, responds to every component of the error, determines the Hamiltonian uniquely, and yields gradients that do not vanish for large orbitalenergy errors. Fine-tuning three Hamiltonian networks on QH9 in this way, which we call FermiScope, reduces the HOMO-LUMO gap error without increasing the element-wise error. The gains carry over to out-of-distribution and molecular-dynamics splits, where the wavefunction-alignment loss helps less.
ER -