arXiv:2609.37733
Foundation Neural-Network Quantum States for Molecular Potential Energy Surfaces in Second Quantization
Lizhong Fu, Jianan Wei, Wenguan Wang, Honghui Shang
Abstract
Second-quantized neural-network quantum states have achieved accurate molecular energies, but extending them across molecular geometries requires a shared representation of the geometry-dependent wavefunction coefficients. We introduce geometry-conditioned foundation neural-network quantum states for molecular electronic structure in second quantization. A single autoregressive model learns a family of ground states from sparse anchor geometries and provides wavefunctions at untrained geometries without further optimization. Orbital alignment matches orbital identities and transports their phases, establishing an aligned orbital basis across geometries. Frozen energies reach chemical accuracy at every untrained query geometry for N$_2$, CO, and H$_4$. On additional molecular paths, the energy-trained wavefunctions yield dipoles, quadrupoles, and natural occupations without property labels. Across three paired N$_2$ 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. At approximately 1 mHa mean absolute error, frozen evaluation reduces the per-geometry cost by $986\times$ relative to independent optimization, yielding an estimated $25.8\times$ end-to-end GPU-cost reduction on a 161-point N$_2$ grid.
AI Summary
The paper addresses whether a single second-quantized neural-network quantum state can represent molecular ground states across geometries, avoiding separate variational optimization at every point on a potential-energy curve. It trains a geometry-conditioned autoregressive Transformer jointly at sparse anchor geometries, then freezes its parameters for queries. Maximum-overlap orbital matching and phase alignment maintain consistent occupation-basis labels; geometry and aligned Fock-diagonal descriptors condition the network. Training minimizes sampled-subspace Rayleigh quotients, while reported energies use full-Hamiltonian local-energy estimators with three independent sampling replicates. Against FCI in the same retained orbital spaces, all untrained queries for N$_2$ and CO in STO-3G and rectangular H$_4$ in cc-pVDZ reportedly fall within $1.6\,\mathrm{mHa}$. Across three paired N$_2$ seeds, orbital alignment reduces query MAE from $34$–$37\,\mathrm{mHa}$ to $0.049$–$0.085\,\mathrm{mHa}$. Energy-trained states also provide dipoles, quadrupoles, and natural occupations on NH$_3$, CO$_2$, and ethylene paths without property labels, although several numerical results remain unresolved placeholders in the supplied text. A separate fixed-budget N$_2$ comparison near $1\,\mathrm{mHa}$ MAE estimates a $986\times$ per-query GPU-cost reduction versus independent optimization. Including training, the 161-point grid costs 1.644 GPU-hours: $25.8\times$ less than independent optimization, $2.53\times$ less than shared pretraining plus fine-tuning, and $1.18\times$ less than sequential warm starts. CPU orbital/integral preparation is excluded. Stated limitations include one-dimensional paths and finite orbital spaces, scaling of sampling and Hamiltonian evaluation, path-dependent orbital transport in multidimensional domains, and unresolved transfer across active spaces or molecules.
The tools also assisted with literature search, manuscript editing, and figure preparation.
Error found: Unresolved numerical placeholders and unavailable plotted values prevent checking the central assertion that every untrained N₂, CO, and H₄ query achieves chemical accuracy. The per-seed ablation table contains unresolved placeholders, leaving the individual paired alignment results unavailable despite numerical ranges being reported in the text.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- 23 pages, 7 figures