Kohn–Sham Hamiltonian learning adds state-space supervision: total-energy error on ∇2DFT falls to one five-hundredth of the strongest baseline
Synopsis
Machine-learned Kohn–Sham Hamiltonians trained only on matrix-element error leave large errors in derived electronic properties, because the operator-to-state mapping is highly anisotropic. DSHam adds supervision on the density matrix, real-space electron density and occupied orbital energies obtained by solving the predicted Hamiltonian. On the ∇2DFT benchmark, electron-density and total-energy errors drop by factors of 40 and 500 relative to the strongest competing method, and the model extrapolates to molecules with twice the training heavy-atom count.
Interpretation
The training target for Hamiltonian predictors expands from matrix-element error alone to the electronic state that the predicted Hamiltonian induces. Such models were typically trained by directly minimizing element-wise mean absolute or squared error between predicted and reference matrices, so supervision stayed in operator space. Across three datasets spanning different molecular scales, models trained with element-wise supervision alone had comparable Hamiltonian errors but markedly different total-energy errors, with errors on larger molecules exceeding chemical accuracy by more than three orders of magnitude.
In training, the predicted Hamiltonian is diagonalized through the generalized eigenvalue problem, and the resulting electronic state is aligned with the reference state. Three added terms cover the density-matrix error in the atomic-orbital basis, the normalized density error on real-space grids, and occupied orbital-energy differences, summed with weights alongside the original Hamiltonian loss. On QH9, the combined three-term loss lowered density-matrix error from 16678.08 to 384.37 and total-energy error from 455.51 to 104.40 (units of 10⁻⁶ and µEh), better than adding any single auxiliary term.
Perspective
The demonstrated scope is three molecular benchmarks, MD17, QH9 and ∇2DFT, plus an extrapolation test trained on QH9 and evaluated on PubChemQC molecules with 10–17 heavy atoms; the molecular-dynamics cases are five flexible molecules of 32–44 atoms from the ∇2DFT test set, run as NVT simulations at 300 K with a 0.5 fs time step.
The method depends on a predefined atomic-orbital basis and does not transfer directly across basis sets; reliable generalization over broader chemical and conformational space remains challenging; Hamiltonian prediction involves higher-order tensor operations and costs more than machine-learning interatomic potentials.