[Paper Review] Machine Learning Hamiltonians are Accurate Energy-Force Predictors

Seongsu Kim, Chanhui Lee+8 more|arXiv (Cornell University)|Feb 18, 2026

Machine Learning in Materials Science0 citations

TL;DR

The paper benchmarks ML Hamiltonians by directly computing energies and forces from predicted Hamiltonians, introducing QHFlow2 which achieves NequIP-level force accuracy and up to 20x lower energy MAE on MD17/rMD17 and QH9 benchmarks.

ABSTRACT

Recently, machine learning Hamiltonian (MLH) models have gained traction as fast approximations of electronic structures such as orbitals and electron densities, while also enabling direct evaluation of energies and forces from their predictions. However, despite their physical grounding, existing Hamiltonian models are evaluated mainly by reconstruction metrics, leaving it unclear how well they perform as energy-force predictors. We address this gap with a benchmark that computes energies and forces directly from predicted Hamiltonians. Within this framework, we propose QHFlow2, a state-of-the-art Hamiltonian model with an SO(2)-equivariant backbone and a two-stage edge update. QHFlow2 achieves $40\%$ lower Hamiltonian error than the previous best model with fewer parameters. Under direct evaluation on MD17/rMD17, it is the first Hamiltonian model to reach NequIP-level force accuracy while achieving up to $20 imes$ lower energy MAE. On QH9, QHFlow2 reduces energy error by up to $20 imes$ compared to MACE. Finally, we demonstrate that QHFlow2 exhibits consistent scaling behavior with respect to model capacity and data, and that improvements in Hamiltonian accuracy effectively translate into more accurate energy and force computations.

Motivation & Objective

• Motivate direct evaluation of ML Hamiltonians as energy–force predictors beyond reconstruction metrics.
• Develop a scalable, accurate MLH model (QHFlow2) with improved robustness for Hamiltonian prediction.
• Establish a unified benchmark to compare Hamiltonian predictors against MLIP baselines using downstream energies and forces.
• Investigate how Hamiltonian accuracy scales with model capacity and data, and how this translates to energy/force accuracy.

Proposed method

• Use equivariant flow matching to map molecular geometry to a Hamiltonian while maintaining rotational equivariance.
• Adopt an SO(2)-equivariant backbone based on eSEN with an efficient two-stage edge update for robustness.
• Construct Hamiltonians as Hermitian blocks via a tensor-expansion readout conditioned on atom types.
• Combine an explicit two-stage pair update to model off-diagonal Hamiltonian blocks and improve robustness to cutoff/radial-basis settings.
• Encode inputs with invariant (Z) and equivariant (H, pair features) representations and predict H in an atom-centered orbital basis.
• Evaluate downstream energies and analytic forces by solving Roothaan–Hall equations with the predicted H and overlap matrix.

Figure 1 : QHFlow2 overall workflow. Given the molecular structure $\mathcal{M}$ , flow time $t$ , and an intermediate Hamiltonian state $\mathbf{H}_{t}$ , QHFlow2 applies an SO(2) backbone, a two-stage pairwise update, and construct Hamiltonian via tensor expansion for energy and force evaluation.

Experimental results

Research questions

• RQ1 Can MLH models predict Hamiltonians with sufficient accuracy to yield accurate energies and forces upon downstream evaluation?
• RQ2 Does QHFlow2's SO(2)-equivariant backbone and two-stage edge updates improve Hamiltonian, energy, and force predictions compared to prior MLH models?
• RQ3 How does Hamiltonian prediction accuracy scale with model capacity and training data, and how does this affect downstream MD quantities?
• RQ4 Do predicted Hamiltonians provide practical benefits such as reduced SCF iterations when used as initial guesses in KS-DFT?

Key findings

• QHFlow2 reduces Hamiltonian MAE by 40–50% relative to prior state of the art with roughly half the parameters.
• On MD17/rMD17, QHFlow2 reaches NequIP-level force accuracy and achieves up to 20x lower energy MAE than NequIP.
• On QH9, QHFlow2 reduces energy error by up to 20x compared to MACE and improves over EquiformerV2.
• QHFlow2 exhibits consistent scaling: increasing model capacity and data size decreases Hamiltonian error and improves downstream energy/force accuracy.
• Predicted Hamiltonians enable reduced SCF iterations in KS-DFT, approaching data-limit references as predictors improve.
• QHFlow2 provides faster inference with better speed-memory trade-offs than prior Hamiltonian models while maintaining or improving accuracy.

Figure 2 : Energy and force accuracy under direct evaluation on the MD benchmark. We report mean absolute errors (MAE) of total energy (top) and forces (bottom) computed from predicted Hamiltonians on six molecular systems. gray bars denote MLIP baselines, green bars denote prior Hamiltonian predict