Can a single neural network predict molecular energies across all geometries without retraining?

Second-quantized neural-network quantum states have delivered accurate molecular energies for years. The bottleneck was extending a trained model across different molecular geometries because each geometry changes the wavefunction coefficients in a way that broke the shared representation.

Researchers from the arXiv submission 2609.37733v1 introduce geometry-conditioned foundation neural-network quantum states. A single autoregressive model learns a family of ground states from sparse anchor geometries. It then provides wavefunctions at untrained geometries without further optimization. The paper spans 23 pages and includes 7 figures.

Orbital alignment keeps orbitals identifiable

Orbital alignment matches orbital identities and transports their phases. This establishes an aligned orbital basis across geometries. Without it, the same orbital index could refer to different physical orbitals at different bond lengths. The alignment step fixes that correspondence.

It works.

The authors test the approach on three molecules. Frozen energies reach chemical accuracy at every untrained query geometry for N₂, CO, and H₄. Chemical accuracy means errors below roughly 1 kcal/mol, a standard benchmark in quantum chemistry.

Error drops by two orders of magnitude

Across three paired N₂ training seeds, orbital alignment lowers the mean absolute energy error over all untrained query geometries from 34-37 mHa to 0.049-0.085 mHa. That is a reduction by a factor of several hundred. The numbers come directly from the abstract.

At approximately 1 mHa mean absolute error, frozen evaluation reduces the per-geometry cost by 986× relative to independent optimization. The same setup yields an estimated 25.8× end-to-end GPU-cost reduction on a 161-point N₂ grid. Independent optimization means training a separate model for each geometry from scratch.

Observables emerge without labels

Energy-trained wavefunctions yield dipoles and quadrupoles without property labels. Natural occupations also appear. The model never sees dipole or quadrupole data during training. Those observables emerge from the wavefunction itself. How well they match reference values is not quantified in the abstract.

The paper covers molecular electronic structure in second quantization. It sits at the intersection of chemical physics, machine learning, and quantum physics. The submission date is 29 September 2026. The authors are Lizhong Fu, Jianan Wei, Wenguan Wang, and Honghui Shang.

Gaps in the public summary

The abstract does not report wall-clock training time for the foundation model. It does not state how many anchor geometries were used for each molecule. It does not compare against traditional quantum chemistry methods like CCSD(T) or FCI on the same grids. It does not discuss transfer to larger molecules or excited states. The full 23-page text may address some of these points.