Transferable deep QMC: usefulness and implementation assessment¶
Assessment date: 2026-11-08. Repository inspected at b42a4e7f.
This is a feasibility assessment, not a solver implementation or numerical
acceptance result. No chemistry calculation or timing experiment was run.
Recommendation¶
Pursue a bounded molecular neural-VMC feasibility experiment if accurate excited-state surfaces are a development priority. This is a substantial new solver programme. It does not directly accelerate the existing SCF, DLPNO or periodic engines. First demonstrate useful accuracy and complete computational cost on a small two-state surface before committing to larger systems.
Primary evidence¶
Schaetzle et al., Nature Computational Science (2026), combine a geometry-conditioned neural wavefunction with parameter sharing across states and energy-dependent directional overlap penalties. The objective includes energy, optional spin and orthogonality terms (equations 1, 6-10); one branch of each overlap uses stopped gradients. Per-geometry moving energy averages determine the penalty direction. The ansatz extends Psiformer (equations 12-22), with nuclear/electronic attention, determinant orbitals, envelopes and a cusp-correcting Jastrow factor.
The reported speedups, approaching 100-fold, compare transferable training with independent deep-QMC calculations, not conventional QC. Demonstrations cover ethylene, C2, methylenimmonium and a 96-electron rubredoxin model. The reported workflow uses NVIDIA A100 GPUs, KFAC and forward-Laplacian automatic differentiation. Transferability concerns geometries of fixed systems, not a universal model across arbitrary compositions.
The supplement, section 1, explicitly reports unsuccessful continuity of near-degenerate C2 singlets. Section 4 discusses parameter-sharing tradeoffs; Table 12 specifies training steps and batches, not portable wall-clock costs. The main text also identifies state-subspace boundary problems and unfinished extensions to varying electron counts. Its coordinate-overlap NAC example does not establish a complete molecular-dynamics workflow.
The full main methods and supplementary discussion sections 1-5 were read; the training table was also inspected visually. The configured local library catalog was searched and contains neither this paper nor its SI. Public publisher copies were available, so no missing-source intake was needed. The earlier arXiv manuscript is a separate version and should not replace the journal paper’s four-system evidence.
Fit with the existing source¶
These are source-level observations, not claims of numerical qualification.
Existing component |
Reusable contribution |
Boundary |
|---|---|---|
Internal comparison methods and state/spin conventions |
Their Gaussian-orbital optimization is different from joint neural-wavefunction training |
|
Cross-geometry overlaps and state-assignment design experience |
Existing small-space, nondegenerate applicability restrictions still apply |
|
A method-independent energy/gradient callback |
Stochastic QMC gradients require their own accuracy and convergence contract |
|
Geometry and output workflow conventions |
A relaxed coordinate scan is not a joint training dataset/sampler |
|
Optional wavefunction exchange and independent-reference workflows |
No verified DeepQMC importer or neural-checkpoint representation is implied |
|
Existing optional-extra and lazy-import conventions |
No neural-VMC/JAX engine was found in the inspected source |
The best application is dense molecular excited-state surfaces with strong correlation, double-excitation character or competing spin states. Transition metal fragments are a longer-term target. For routine single geometries, ground-state optimizations or current periodic performance work, this assessment establishes no advantage. Periodic extension would require its own boundary conditions, Coulomb treatment, ansatz and validation.
Proposed implementation boundaries¶
Implement scientific kernels within vibe-qc. The repository policy excludes
importing another QC program inside python/vibeqc/ or cpp/; an internal
DeepQMC wrapper would violate that rule. Use DeepQMC only as an independently
invoked optional reference in tests/examples. Its MIT-licensed source
can inform development; any adapted code needs retained notices and an entry
in the licence documentation. Dependency and copied-file terms must be checked
at the pinned version. JAX and an appropriate optimizer are numerical tools;
introduce them through optional extras with lazy imports if selected.
A proposed internal vibeqc.qmc package would own these responsibilities:
Signed/logarithmic wavefunction evaluation, stable determinants, electron permutations, cusp conditions and differentiable kinetic local energy. Include nuclear repulsion consistently in reported molecular energies.
Reproducible Metropolis sampling, per-geometry/state chains, re-equilibration, autocorrelation-aware uncertainties and checkpoints that retain random state and sampler state.
Ground-state VMC optimization and independent post-training evaluation. A CPU diagnostic path is useful, but performance qualification needs suitable accelerator hardware and measured memory/precision behaviour.
Multistate energy/overlap/spin estimators and the directional objective. Preserve the paper’s stop-gradient semantics and adaptive penalty scaling; replacing these with a symmetric overlap penalty is a material departure.
Geometry-conditioned attention and envelopes, parameter sharing and per-geometry ordering. Averaging parameters after per-state updates, as in the paper, is not automatically equivalent to a single jointly optimized shared parameter tree under a preconditioned optimizer. Test equivalence or document the departure.
Results routed through
vibeqc.output, including statistical uncertainty, state character, overlap, spin, Hamiltonian identity and training metadata. Neural checkpoints need an explicit format rather than an assumed TREXIO mapping. Cite the algorithm beside the implementation and register its user-visible citation in the citation database.
Keep forces, molecular dynamics, ECPs and local pseudohamiltonians outside the first experiment. Existing Gaussian ECP matrix elements are not a real-space QMC nonlocal-potential evaluator. Introduce those capabilities separately with matched Hamiltonians and reference evidence.
Proposed milestones and acceptance¶
The coarse roadmap and GitLab parent #995 track this research programme. Bounded stages are #996 (ground-state engine), #997 (multistate constraints), #998 (transferable training), #999 (ethylene reproduction and complete costs) and #1000 (C2 robustness). Each issue includes literature attachments, dependencies and acceptance criteria. These are planned tasks, not completed implementation or numerical evidence.
First: ground-state engine. Use small all-electron H2/He cases, with analytic checks where applicable and converged independent numerical references. Establish the energy estimator, sampling errors and repeatability. This is an infrastructure milestone, not a reproduction of the paper.
Second: the bounded scientific pilot. Compute two singlet ethylene surfaces using the paper’s geometries. Compare independent training, transfer across geometries, shared state parameters and dynamic ordering as separate ablations. Compare with the published dataset and independently run pinned DeepQMC; use vibe-qc multireference results as additional diagnostics rather than exact truth. Include held-out geometries to measure interpolation.
Submit complete calculations through vq. Record actual source/native and
numerical-backend identities, seeds, hardware, precision, memory, training
and evaluation cost. Choose numerical tolerances before evaluating outcomes.
Report excitation gaps, surface-shape errors, overlaps/state character, spin
expectations and uncertainty across independent seeds. Reproduce the paper’s
single-global-offset relative-energy metric (equation 23); do not independently
shift each state or geometry, which would conceal relative-energy errors.
Measure full accelerator-hours and elapsed cost at matched accuracy, including compilation, equilibration and final evaluation. Iteration-count ratios alone cannot establish a practical speedup. Compare against both independent neural VMC and a relevant conventional method.
Third: robustness. Add C2 as an adversarial near-degeneracy test, including the documented continuity failure. State labels need overlap/subspace diagnostics; sorting energies alone is insufficient. Mark ambiguous state assignments explicitly. Add larger geometry grids only after the pilot shows benefit. ECPs, transition metals and gradients each need separate acceptance.
Decision: feasible and scientifically worthwhile as an experimental molecular solver programme; performance, reliable state continuity and general-purpose photodynamics remain unverified for vibe-qc. There is insufficient measured evidence for a delivery-time estimate or a promised speedup.