acceptodds
Withdrawn from ICLR 2027

The NNP Score Completeness Boundary: Physics Decomposition for Crystal Structure Prediction with Neural Network Potentials

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Crystal structure prediction (CSP) faces a trade-off between data-driven generative models, which need substantial training data, and physics-based sampling, which is computationally expensive. Pre-trained neural network interatomic potentials (NNPs) offer a middle ground: their energy gradients define conservative scores, so they can in principle sample the Boltzmann distribution without training a generative model. We show that using them directly exposes a fundamental limitation. NNPs are not uniformly complete: they capture long-range chemistry and equilibrium forces accurately, but their training data contain no deep-overlap configurations, leaving short-range Pauli repulsion unconstrained and driving noise-dependent trajectories into an out-of-distribution (OOD) regime. We introduce a physics decomposition that pairs learned long-range chemistry with analytically specified short-range universality. The augmented score is exactly conservative and needs no generative training. Experiments show that the augmented score holds the induced rate to at most a few percent across the entire scan. The frontier where the score stops being complete—the NNP score completeness boundary—is a property of the training data that any machine-learned potential inherits, not an architectural defect. Supplying the missing short-range physics analytically is what makes a pre-trained potential a usable generative score. Our code is available at https://anonymous.4open.science/r/MaterialGen_anoy-3A8B.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
crystal structure generation, neural network interatomic potentials, score completeness boundary

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_7qjKOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper uses a pretrained neural network potential (NNP) as a training-free score inside annealed Langevin or probability-flow ODE sampling for crystals. It argues that NNPs trained on relaxed structures have no data for very close atom pairs, so noisy sampling drives atoms into overlap where the learned force is unreliable. The authors define a “completeness boundary” at 0.75 times the sum of covalent radii and add three protection layers: 1) an analytic ZBL-matched repulsion below the boundary 2) a density-scaled noise 3) a hard constraint that projects close pairs back onto the boundary. Each run starts from a known MP-20 test structure plus Gaussian noise and anneals it back down, which the authors frame as the relaxation stage of crystal structure prediction. A noise scan with MACE-MP-0 shows the bare sampler fails above a noise threshold, while the protected sampler keeps the failure rate near zero. A benchmark with eSEN reports validity, stability, SUN and match rate. My expertise is diffusion and score-based sampling rather than interatomic potentials, so I focus on the method and the evaluation.

Critical strengths
  1. Clear and practical observation. Pretrained potentials are not trained on overlapping configurations, so using them directly as a sampling score fails at high noise. The paper measures this with large trajectory counts across 20 compositions and two potentials. The bare failure curve is a useful reference for anyone using an NNP inside a sampler.
  2. Careful and honest reporting. The authors separate start-time failures from sampler-induced ones, pair the layer ablation, and check threshold-free statistics. They state openly which protocols were not run and where their theorems do not apply.
Critical weaknesses
  1. The method is a set of protection layers for relaxation, and the layers are not new. Each run starts from the known test structure plus noise, so it recovers a given structure rather than generating from a composition. The composition-only start was not run. Much of the failure comes from injecting up to several Å of noise into a relaxation, which standard relaxation does not do. The layers themselves are standard: ZBL repulsion spliced into ML potentials is common and already built into MACE-MP-0, and a minimum-distance constraint is a standard device in structure search. The paired ablation shows the hard constraint (L3) does almost all the work, density-scaled noise adds almost nothing, and the step-rejection guard is harmful. The repulsion-off arm that would show whether the analytic wall adds anything beyond the potential’s own repulsion was not run. The boundary itself is a chosen threshold at the edge of the training data, and where the potential’s forces actually become wrong is never measured.
  2. Downstream effectiveness and theory are not connected to the claims. The OOD rate and validity both test minimum interatomic distance, which L3 enforces directly, so its win on these metrics is close to circular. Stability uses NNP energies on a hull restricted to MP-20 phases with no DFT check, and SUN decreases with protection on PF-ODE. There are no baselines at matched settings: no plain FIRE relaxation of the noised start, no MD annealing or basin hopping, and no learned generator. Many of the described protocols, including DFT validation, efficiency, cross-potential and score comparison, are not executed. The theory is mostly general and does not explain why the layers work. Conservativity holds for any energy gradient. The steady-state result does not cover density-scaled noise. The reflection theorem is stated for an idealized process rather than the implemented projection. The reliability condition is refuted by the authors’ own measurements.
Questions for authors
  1. Compare against 1) FIRE or BFGS relaxation of the same noised starts 2) MD annealing and basin hopping at matched force evaluation 3) a learned generator conditioned on the same compositions.
  2. Report the composition-only start, so the method is evaluated as generation rather than recovery of a given structure. Also report DFT checks on a subset of the stable candidates.
  3. Run 1) the repulsion-off arm at matched drift caps and 2) a DFT force scan across the threshold, to show that the boundary and the analytic wall matter beyond the hard constraint.
Reviewer_ke8xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on the usage of pre-trained neural network for potential energy gradient to provide the score of a Boltzmann distribution such that one can sample from the Boltzmann distribution via Langevin Dynamics. However, annealing needs relatively large noise, which can lead to atomic overlap, a situation that has never been seen during training of these models, leading to OOD gradients. They propose to add an analytic short-range repulsion plus a projection to separate overlapping atoms at generation time.

Critical strengths
  • The problem is well presented/derived and also shown.
  • The derivation of their solution is intuitive and sensible.
  • The paper is very open and clear about its limitations/shortcomings.
Critical weaknesses
  • Not sure if this really is that relevant to ICLR, the unreliablility of the score on OOD is well known, the rest is somewhat heuristic fixes which are specific to Material sciences.
  • The proposed fix seem not to be the main improvement. The ablation in table 9 shows the majority of improvement to come from the projection (L3) and not the repulsion term.
  • The experiments measure recovery not generation.
  • Missing baselines, e.g. other generative models, MD, Basin Hopping, simply relaxing the noised structure with a standard optimizer on the same potential.
Questions for authors

See weaknesses

Reviewer_nAXKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors argue that pre-trained MLP models fail to generate realistic predictions on chemical properties of crystal structures when the parameters go out-of-distribution particularly with the introduction of noise and short-range pauli repulsion effects at short distances. They mitigate this by adding an analytic repulsion constraint and a min distance constraint. Experiments on energy relaxation corroborate their findings.

Critical strengths

The authors try to tackle a well-known and difficult problem in the MLP (machine learning potentials' space) is OOD generalization of pre-trained and regular trained MLP models.

I appreciate the careful control of the noise level as a knob to continuously control the strength of the induced OOD-ness which also benchmarks the variation of the strength of various mitigation strategies.

Critical weaknesses

The writing is quite poor and impenetrable due its high info density.

The main concern is the limited novelty since adding coulomb repulsion terms to potentials is well known and enforcing min clamps is also well known. These are band-aid fixes to the generalizability problem of MLP functions on data regimes that are outside the ones covering the training sets.

Lots of crucial baselines are missing e.g. standard second-order gradient descent methods like L-BFGS or other ML methods. Computing DFT potentials for the studied settings is also missing.

There might be a circular issue in the evaluation and mitigation strategies. From what I can tell, eval counts a failure if interatomic distance r_ij is below some threshold but the fix projects it to a larger distance anyway.

Questions for authors

see above.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.