Pith. sign in

REVIEW 3 major objections 4 minor 71 references

Atom-averaged hidden features of a pretrained interatomic potential can evaluate crystal generators for quality and memorization — and can guide the generation itself.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 00:23 UTC pith:PHIAPRNO

load-bearing objection CFTD is a smart dual-featurizer metric that fixes a real TNovD tradeoff, but its MACE-based quality proxy is validated on in-distribution data only, and circularity affects the generator evaluation — worth refereeing, not canonizing yet. the 3 major comments →

arxiv 2607.28776 v1 pith:PHIAPRNO submitted 2026-07-30 cs.LG cond-mat.mtrl-sci

Representations from Pretrained Machine-Learning Interatomic Potentials as Coarse Coordinates for Material Generation and Evaluation

Paul Hagemann , Katharina Ueltzen , Simon M\"uller , Janine George , Philipp Benner This is my paper
classification cs.LG cond-mat.mtrl-sci
keywords Coarse-Fine Transport Distancecrystal structure generationmachine learning interatomic potentialsMACE featuresgenerative model evaluationoptimal transportmemorization detectionflow matching
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that evaluating a crystal-structure generator well requires two separate notions of closeness: a fine-grained identity measure that detects memorized copies of training structures, and a coarse-grained quality measure that captures whether a structure is physically plausible, including implicit stability. It packages both into a single distributional distance, the Coarse-Fine Transport Distance (CFTD), built on an optimal-transport plan between training and generated sets. The coarse quality term uses mean-pooled, randomly projected hidden features from a pretrained machine-learning interatomic potential (MACE); the fine identity term uses a contrastive GNN trained with augmentations that include chemically similar substitutions. The paper shows CFTD flags mode collapse and overfitting that instance-level metrics miss, and it demonstrates that the same coarse MACE features can condition a flow-matching generator, improving validity and stability on a standard benchmark while increasing memorization.

Core claim

The central discovery is that hidden, invariant features from a pretrained MACE interatomic-potential model — averaged over atoms and compressed by a fixed random projection — form a 'coarse coordinate' space that carries meaningful information about crystal-structure quality, despite not being trained for that purpose. Embedding these coordinates in the CFTD optimal-transport distance, alongside a contrastive identity featurizer, yields a metric that penalizes generated crystals that are either copies of training materials (too close in identity space) or outside the quality distribution (too far in MACE space), without forcing a tradeoff between the two. The paper demonstrates this on pert

What carries the argument

The central object is the Coarse-Fine Transport Distance (CFTD), an optimal-transport distributional metric between the empirical training distribution and the empirical generated distribution. It uses two featurizers: a 'fine' identity featurizer, a contrastive GNN (InfoNCE loss) whose augmentation function adds Gaussian coordinate noise and substitutions with chemically similar elements (modified Pettifor neighbors), and a 'coarse' quality featurizer, obtained by mean-pooling the invariant hidden node features of a pretrained MACE interatomic potential across atoms, concatenating layers, and applying a fixed Gaussian random projection. A single OT plan is computed with a cost that averages

Load-bearing premise

The quality branch of CFTD assumes that Euclidean distances between mean-pooled, randomly projected MACE hidden features are a reliable proxy for physical quality or implicit stability, including for generated crystals outside the training distribution.

What would settle it

Take a generated structure that survives a full MACE or DFT relaxation with low energy above hull but lies far from all training structures in the projected MACE feature space; if CFTD labels it low-quality while direct stability measures call it stable, the quality proxy fails. A cheaper check: on a larger polymorph set than SiO2 and ZnS, count cases where MACE feature distance and hull-energy difference strongly disagree.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • CFTD can be applied before and after MLIP relaxation; relaxed structures often reveal memorization that is invisible in unrelaxed coordinates, making the metric a practical overfitting checkpointer during generator training.
  • Because CFTD is distributional, it penalizes mode-collapsed generators even when their individual outputs are stable, unique, and novel — a failure mode that instance-level continuous SUN metrics can miss.
  • On a benchmark of published crystal generators, the metric consistently ranks the held-out test set and one leading generative model highest, while a reinforcement-learning-biased model ranks lowest, matching the hypothesis of overconcentration.
  • Feeding coarse MACE features as a conditioning signal to a flow-matching generator improves validity and stability on a standard benchmark, confirming that the same representation used for evaluation carries a usable generative signal.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the coarse MACE feature space is itself learnable, a hierarchical generator that first samples new feature vectors from a learned prior could combine the stability gains observed here with higher novelty, effectively creating a latent-space material design tool.
  • The identity featurizer's augmentation set is pluggable; the same CFTD scaffold could encode different definitions of 'same material' (e.g., ignoring substitutions, or treating them as new), letting a lab tune novelty to its use case without changing the rest of the metric.
  • Because MACE feature distances correlate with polymorph hull-energy differences, a MACE-conditioned generator could be steered toward low-energy regions of feature space, yielding a stability proxy for conditional generation without requiring DFT labels at generation time.
  • The paper leaves open a head-to-head comparison of CFTD with validation loss as a checkpoint criterion; one testable extension is whether CFTD-based early stopping yields generators with better downstream stability than loss-based early stopping.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper proposes CFTD, a distributional metric for evaluating generative models of crystals. It combines a 'fine' contrastive-GNN featurizer that treats chemically similar or perturbed structures as identical, and a 'coarse' featurizer based on mean-pooled, randomly projected MACE hidden features, assumed to encode implicit stability. The metric computes an optimal-transport plan between training and generated sets using the average of the two distance matrices, then penalizes pairs that are too close in the identity featurizer (memorization) or too far in the MACE featurizer (low quality), with thresholds calibrated on a validation set. The paper validates CFTD on perturbed MP20 and Perov-5 data, compares it with continuous SUN metrics, benchmarks nine generative models before and after relaxation, and introduces a flow-matching generator conditioned on empirical MACE features, reporting improved LeMat-GenBench validity and stability but lower novelty.

Significance. The dual-featurizer decomposition of quality versus memorization is a conceptually sensible response to the known difficulty of separating these axes with a single featurizer. The theoretical bounds in Appendix B are broadly correct under the stated assumptions and give a formal sense in which low CFTD controls both quality coverage and novelty. The toy experiments in Fig. 2 demonstrate qualitative sensitivity to perturbations and leakage, and the code release with a pre-trained identity featurizer is a practical contribution. If the MACE-distance stability proxy survives out-of-distribution validation, CFTD would be a useful complement to instance-level SUN metrics. The main open risks are the under-validated implicit-stability proxy and the partially circular MACE-conditioning experiment; both are addressable with additional empirical work.

major comments (3)
  1. [§2, Algorithm 1; Appendix F] The quality penalty in CFTD is ReLU(C^{mace}_{i,j} - τ_qual), so the quality term is entirely governed by Euclidean distances between mean-pooled, randomly projected MACE features. This proxy is validated only in-distribution: Appendix F shows an MLP can predict MP20 formation energies (Fig. 11) and a correlation for SiO2 and ZnS polymorphs with visible outliers (Fig. 12). This does not establish monotonicity with physical stability for out-of-distribution generated structures such as Chemeleon2 samples, which the paper ranks lowest. If the proxy fails OOD, the central ranking claims could be artifacts of the feature-space geometry. Please add an independent validation on generated samples, e.g., compute energy-above-hull or another non-MACE stability measure for models that CFTD ranks high and low, and test whether the CFTD quality term co-ranks with it.
  2. [§4.1, Fig. 6] The conclusion that MACE conditioning improves stability is partially circular in the CFTD evaluation. The conditioning signal is the empirical MACE feature distribution of the training set, and the CFTD quality component uses the same MACE featurizer; generated samples conditioned on train-set MACE features are therefore placed in the low-penalty quality regime by construction. The paper acknowledges this in §4.1, but the LeMat-GenBench numbers (V=98.5 vs 97.2, S=11.7 vs 7.7) are modest point estimates and are not backed by uncertainty or significance. An ablation conditioning on random features, or an independent stability metric on the generated samples, is needed before claiming that MACE features genuinely guide generation toward stability.
  3. [§3, §3.1, Figs. 2–6] All CFTD comparisons are reported as point estimates from single evaluation sets, with no error bars, confidence intervals, or significance tests. Given the finite-sample OT coupling and the heuristic calibration of τ_mem, τ_qual, and M, the ranking differences (e.g., MatterGen vs Crystalite, Chemeleon2 last) may not be robust. Appendix I indeed shows that changing M changes the relaxed-model ranking: with M_new = 0.5M, Crystalite becomes the best model. Please report variability across model seeds or evaluation subsets and state which ranking differences are significant.
minor comments (4)
  1. [§2, MACE Featurizer paragraph] The text says the random projection 'allows us to reduce the dimensionality of the pooled MACE features, while keeping all information intact.' This is not mathematically accurate when n_proj < dim(H), and Appendix E.1 itself shows that reconstruction accuracy improves with projection length. Please rephrase and discuss the lossy nature of the projection and the choice of n_proj.
  2. [Fig. 2 caption / §3] The text says 'The results in Fig.2 prove that problematic structures and data leakage are indeed punished.' For toy experiments, 'demonstrate' or 'indicate' would be more appropriate than 'prove.'
  3. [Appendix G] The Perov-5 experiments reuse the Identity featurizer trained on MP20. This is reasonable as a transfer test, but the text should state that the fine featurizer was not retrained for Perov-5 and discuss the possible effect on the memorization component.
  4. [Table 1] The CFTD ranking is reported for β=0.6 and calibrated M. It would be helpful to state explicitly that other β values in Appendix H give qualitatively similar rankings, while M variations in Appendix I do change some relaxed-model rankings.

Circularity Check

1 steps flagged

CFTD's quality assessment of the MACE-conditioned generator is partly by construction, since the generator is conditioned on the same MACE feature space that defines CFTD's quality penalty.

specific steps
  1. self definitional [Sec. 4.1 (Evaluation; Fig. 6), Algorithm 1, Sec. 2 ('The Coarse in CFTD')]
    "we use the empirical distribution of the MP20 train set as the prior over these conditions. ... The cosine distances in Fig. 6 (b) indicate a very high degree of agreement between the features of the generated materials and the conditions ... Note that while the CFTD evaluation of the MACE-conditioned model shares the same underlying MACE representation, the pre-relaxation and evaluation procedure of LeMat-GenBench does use a more thorough relaxation procedure with 3 different MLIPs."

    Algorithm 1 defines the quality penalty as ReLU(C^mace_{i,j} - tau_qual) with C^mace_{i,j} = ||F_mace(x_i) - F_mace(g_j)||, where F_mace is the mean-pooled, random-projected MACE-MP-0b3 feature map. The MACE-conditioned generator draws its condition h from the empirical distribution of training-set MACE features and is trained to reproduce that condition; Fig. 6b confirms the generated materials' F_mace features closely match the conditions. Consequently, generated samples are close to training samples in exactly the F_mace geometry used by the quality term, so the ReLU quality penalty is suppressed by construction. The paper's own note concedes the shared representation. The observed 'better quality coverage' of the MACE-conditioned model is therefore partly an identity check rather than

full rationale

The paper's core metric construction is mostly self-contained: CFTD is defined in Algorithm 1 from two featurizers, thresholds are calibrated on train/validation splits, and the metric is probed with perturbation/leakage toy experiments and compared against cSUN. The identity featurizer is described and trained; no load-bearing uniqueness theorem or ansatz is imported via self-citation. The central circularity is confined to Sec. 4.1: the MACE-conditioned generator is conditioned on the empirical F_mace values of training structures, and the CFTD quality term is the distance in that same F_mace space to training structures. Fig. 6b confirms the generator reproduces the conditioning features, so the low quality penalty is partly by construction. The paper explicitly acknowledges the shared representation and points to LeMat-GenBench as independent evidence; that external benchmark does support the generator benefit, so the paper does not reduce entirely to its own metric. The quality proxy itself (Appendix F) is validated only in-distribution and with outliers; that is a robustness/correctness concern, not circularity. Similarly, using MACE both to relax structures and to score quality conflates the two uses, but this is not a formal equation-level reduction. Overall, one 'prediction' reduces by construction while the metric's general framework and external benchmark retain independent content, giving partial circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central metric rests on a few free hyperparameters (τ, M, β, projection dimension, augmentation choices) and on two empirical proxies: MACE features for quality and contrastive features for identity. No new physical entities are introduced. The main circularity risk is that the quality featurizer is simultaneously used for evaluation and for conditioning the generator; the paper acknowledges this but does not fully decouple them.

free parameters (5)
  • τ_mem and τ_qual (Goldilocks thresholds) = not reported numerically; chosen as quantiles of coupled train-validation mass (1-β)/2
    These thresholds define the memorization and low-quality regimes. They are fitted to the calibration split and depend on β; different choices would change CFTD values.
  • M (balance factor) = calibrated to balance memorization and quality penalties on validation set
    M scales the memorization penalty relative to quality. Ablation in Appendix I shows M changes relaxed-model rankings, so it is a genuine free knob.
  • β (Goldilocks zone proportion) = default 0.6; ablated over 0.4, 0.6, 0.8
    The fraction of mass left unpenalized. The paper ablaties it and shows rankings are mostly stable, but the value itself is a design choice.
  • Random projection dimension n_proj = not stated for CFTD; reconstruction experiments use 32, 64, 128, 256
    The dimension of the coarse MACE feature space after random projection is a free choice; larger projections encode more information (Appendix E).
  • Contrastive augmentation hyperparameters = Gaussian noise std 0.01, Pettifor substitution probs 0.1/0.05 (main text); training uses 0.3/0.01, noise prob 0.5, scale
    These define the identity featurizer's notion of 'same material.' They are hand-chosen and directly control what is counted as memorization.
axioms (5)
  • standard math Optimal transport with a weighted-average cost matrix C = 0.5*C_id + 0.5*C_mace is a sensible way to couple training and generated samples.
    The metric's construction in Algorithm 1 relies on OT as a distributional discrepancy; this is standard but the specific cost averaging is a modeling choice.
  • domain assumption Mean-pooled, random-projected MACE hidden features are a reliable implicit-stability proxy for the quality of generated structures.
    The paper supports this with formation-energy regression and polymorph correlations (Appendix F), but outliers and out-of-distribution behavior remain; this is the weakest load-bearing assumption.
  • domain assumption The identity featurizer trained with InfoNCE and the chosen augmentations defines a chemically meaningful notion of 'same material' (memorization).
    The contrastive GNN is trained on MP20 with Gaussian noise and Pettifor substitutions; its detection of substitutes depends on these augmentations.
  • domain assumption The non-degeneracy condition in Theorem B.2: if a generated sample is r_mem-close to a training sample in identity space, the OT plan assigns it at least α/m mass.
    Required for the novelty bound; it is stated as an assumption and not verified numerically, so the theoretical guarantee may not hold in practice.
  • domain assumption The generated structures from refs. [12] and [7] are representative outputs of the underlying models.
    Benchmarking relies on externally provided generations; if those sets are cherry-picked or nonstandard, the rankings could shift.

pith-pipeline@v1.3.0-alltime-deepseek · 25725 in / 11284 out tokens · 117928 ms · 2026-08-03T00:23:24.447947+00:00 · methodology

0 comments
read the original abstract

Generative machine learning is increasingly used for inorganic crystal structure generation. Most models and the corresponding evaluation approaches rely on simple forms of crystal structure representation. In this paper, we showcase the power of atom-averaged features from pretrained Machine-Learning Interatomic Potentials (MLIPs), such as MACE, for such tasks. We first introduce a distance measure that assesses the output of material generative models by capturing both quality and novelty in a single distribution-based evaluation framework. In particular, we introduce the Coarse-Fine Transport Distance (CFTD) using two different featurizers, where the quality component is based on coarse MACE features. We showcase CFTD's versatility in capturing crystal-structure quality while also detecting memorization, and compare it with the recently introduced continuous SUN metrics. We further show that coarse MACE features can be used as guidance for a material generative model.

Figures

Figures reproduced from arXiv: 2607.28776 by Janine George, Katharina Ueltzen, Paul Hagemann, Philipp Benner, Simon M\"uller.

Figure 1
Figure 1. Figure 1: Overview of the CFTD metric and the generative pipeline. We propose using mean [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: CFTD behavior of the MP20 train and test set across different perturbation settings of the [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of continuous SUN versus CFTD. [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of common material generative models: relaxed and unrelaxed. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of our proposed material generator trained on an MP20 subset of 2,000 ma [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CFTD of models. Cosine similarities are for materials generated conditioned on empirical [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Composition and lattice reconstruction curves on MP20. [PITH_FULL_IMAGE:figures/full_fig_p023_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Space group reconstruction curves. Regarding the space group reconstruction, we report the accuracy of the predicted top space group for the two datasets in Fig.8. The space group accuracy is around 0.35-0.40 on the MP20 and 0.55- 0.58 on the Carbon24 dataset. These experiments show that the projections of the mean-pooled MACE features capture both chem￾ical and geometric information. However, they do not … view at source ↗
Figure 9
Figure 9. Figure 9: Separation between the MACE featurizer and the Identity featurizer. [PITH_FULL_IMAGE:figures/full_fig_p024_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Joint histograms of cosine similarities between random pairs of clean or noisy data for [PITH_FULL_IMAGE:figures/full_fig_p024_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: MAE of formation energy per atom over epochs for different lengths of mean pooled [PITH_FULL_IMAGE:figures/full_fig_p025_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Polymorph plots showing hull-energy differences as a function of MACE feature L2 [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: CFTD behavior across different perturbation settings on Perov-5. [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of common material generative models: relaxed and unrelaxed for varying [PITH_FULL_IMAGE:figures/full_fig_p028_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Comparison of common material generative models: relaxed and unrelaxed for varying [PITH_FULL_IMAGE:figures/full_fig_p029_15.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

71 extracted references · 6 linked inside Pith

  1. [1]

    Pierre-Paul De Breuck, Hai-Chen Wang, Gian-Marco Rignanese, Silvana Botti, and Miguel A. L. Marques. Generative AI for crystal structures: A review.npj Computational Materials, 11(1):370, December 2025

    2025

  2. [2]

    A generative model for inorganic materials design.Nature, 2025

    Claudio Zeni, Robert Pinsler, Daniel Z ¨ugner, Andrew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabb ´e, Shoko Ueda, Roberto Sordillo, Lixin Sun, Jake Smith, Bichlien Nguyen, Hannes Schulz, Sarah Lewis, Chin-Wei Huang, Ziheng Lu, Yichi Zhou, Han Yang, Hongxia Hao, Jielan Li, Chunlei Yang, Wenjie Li, Ryota Tomioka, and ...

    2025

  3. [3]

    Jaakkola

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi S. Jaakkola. Crystal diffusion variational autoencoder for periodic material generation. InInternational Conference on Learning Representations, 2022

    2022

  4. [4]

    DMFlow: Disordered materials generation by flow matching

    Liming Wu, Qi Li, Mingze Li, Songyou Li, Shifeng Jin, and Wenbing Huang. DMFlow: Disordered materials generation by flow matching. 2026

    2026

  5. [5]

    Andersson, Abhijith S Parackal, Dong Qian, Rickard Armiento, and Fredrik Lindsten

    Filip Ekstr ¨om Kelvinius, Oskar B. Andersson, Abhijith S Parackal, Dong Qian, Rickard Armiento, and Fredrik Lindsten. WyckoffDiff – a generative diffusion model for crystal sym- metry. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd Internationa...

    2025

  6. [6]

    Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset.Mater

    Mikkel Juelsholt. Continued challenges in high-throughput materials predictions: MatterGen predicts compounds from the training dataset.Mater. Horiz., 13(11):5672–5679, June 2026

    2026

  7. [7]

    Gleason, Ali Ramlaoui, Andy Xu, Georgia Channing, Daniel Levy, Cl ´ementine Fourrier, Nikita Kazeev, Chaitanya K

    Siddharth Betala, Samuel P. Gleason, Ali Ramlaoui, Andy Xu, Georgia Channing, Daniel Levy, Cl ´ementine Fourrier, Nikita Kazeev, Chaitanya K. Joshi, S ´ekou-Oumar Kaba, F ´elix Therrien, Alex Hernandez-Garcia, Roc ´ıo Mercado, N. M. Anoop Krishnan, and Alexandre Duval. Lemat-genbench: A unified evaluation framework for crystal generative models.arXiv prep...

    arXiv 2026

  8. [8]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochre- iter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. V on Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associa...

    2017

  9. [9]

    Elena, D ´avid P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, D ´avid P. Kov´acs, Janosh Riebe- sell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Filippo Bigi, Samuel M. Blau, Vlad C ˘arare, Michele Ce- riotti, Sanggyu Chong, James P. Darby, Sandip De, Flaviano Della Pia, V olker L. Deri...

    2025

  10. [10]

    Transport novelty dis- tance: A distributional metric for evaluating material generative models.arXiv preprint arXiv:2512.09514, 2025

    Paul Hagemann, Simon M ¨uller, Janine George, and Philipp Benner. Transport novelty dis- tance: A distributional metric for evaluating material generative models.arXiv preprint arXiv:2512.09514, 2025

    arXiv 2025

  11. [11]

    Computational optimal transport.arXiv preprint arXiv:1803.00567, 2020

    Gabriel Peyr ´e and Marco Cuturi. Computational optimal transport.arXiv preprint arXiv:1803.00567, 2020

    Pith/arXiv arXiv 2020

  12. [12]

    Continuous SUN (stable, unique, and novel) metric for generative modeling of inorganic crystals.Machine Learning: Science and Technology, 7(3):035064, June 2026

    Masahiro Negishi, Hyunsoo Park, Kinga Oliwia Mastej, and Aron Walsh. Continuous SUN (stable, unique, and novel) metric for generative modeling of inorganic crystals.Machine Learning: Science and Technology, 7(3):035064, June 2026

    2026

  13. [13]

    Crystal structure prediction by joint equivariant diffusion

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal structure prediction by joint equivariant diffusion. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  14. [14]

    Space group constrained crystal generation

    Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group constrained crystal generation. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  15. [15]

    Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary Ward Ulissi

    Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary Ward Ulissi. All-atom diffusion transformers: Unified gen- erative modelling of molecules and materials. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, edi- tors,Pro...

    2025

  16. [16]

    Exploration of crystal chemical space using text-guided generative artificial intelligence.Nat Commun, 16, 2025

    Hyunsoo Park, Anthony Onwuli, and Aron Walsh. Exploration of crystal chemical space using text-guided generative artificial intelligence.Nat Commun, 16, 2025

    2025

  17. [17]

    Guiding generative models to uncover diverse and novel crystals via reinforcement learning.arXiv preprint arXiv:2511.07158, 2025

    Hyunsoo Park and Aron Walsh. Guiding generative models to uncover diverse and novel crystals via reinforcement learning.arXiv preprint arXiv:2511.07158, 2025

    arXiv 2025

  18. [18]

    Revisiting precision recall definition for genera- tive modeling

    Loic Simon, Ryan Webster, and Julien Rabin. Revisiting precision recall definition for genera- tive modeling. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 5799–5808. PMLR, 09–15 Jun 2019

    2019

  19. [19]

    Precision recall cover: A method for assessing generative models

    Fasil Cheema and Ruth Urner. Precision recall cover: A method for assessing generative models. In Francisco Ruiz, Jennifer Dy, and Jan-Willem van de Meent, editors,Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, volume 206 of Proceedings of Machine Learning Research, pages 6571–6594. PMLR, 25–27 Apr 2023

    2023

  20. [20]

    The unreason- able effectiveness of deep features as a perceptual metric

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreason- able effectiveness of deep features as a perceptual metric. InCVPR, 2018

    2018

  21. [21]

    Mastej, and Aron Walsh

    Masahiro Negishi, Hyunsoo Park, Kinga O. Mastej, and Aron Walsh. Continuous unique- ness and novelty metrics for generative modeling of inorganic crystals. InAI for Accelerated Materials Design - NeurIPS 2025, 2025. 13

    2025

  22. [22]

    Resolving the data ambiguity for periodic crystals

    Daniel Widdowson and Vitaliy Kurlin. Resolving the data ambiguity for periodic crystals. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors,Advances in Neural Information Processing Systems, volume 35, pages 24625–24638. Curran Associates, Inc., 2022

    2022

  23. [23]

    Substitution-Based Analysis of Structural Novelty for Generative Models of Materials.arXiv preprint arXiv:2606.23166, (arXiv:2606.23166), June 2026

    Masahiro Negishi and Aron Walsh. Substitution-Based Analysis of Structural Novelty for Generative Models of Materials.arXiv preprint arXiv:2606.23166, (arXiv:2606.23166), June 2026

    Pith/arXiv arXiv 2026

  24. [24]

    The unification of representation learning and generative modelling, 2025

    Kieran Didi. The unification of representation learning and generative modelling, 2025

    2025

  25. [25]

    Maxime Oquab, Timoth ´ee Darcet, Th´eo Moutakanni, Huy V . V o, Marc Szafraniec, Vasil Khali- dov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patr...

    2024

  26. [26]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022

    2022

  27. [27]

    Equivariant diffusion for structure-based de novo ligand generation with latent-conditioning.Journal of Cheminformat- ics, 17(1):90, 2025

    Tuan Le, Julian Cremer, Djork-Arne Clevert, and Kristof T Sch ¨utt. Equivariant diffusion for structure-based de novo ligand generation with latent-conditioning.Journal of Cheminformat- ics, 17(1):90, 2025

    2025

  28. [28]

    Geometric representation condition improves equivariant molecule generation

    Zian Li, Cai Zhou, Xiyuan Wang, Xingang Peng, and Muhan Zhang. Geometric representation condition improves equivariant molecule generation. InForty-second International Conference on Machine Learning, 2025

    2025

  29. [29]

    Platonic representation of foundation machine learning inter- atomic potentials.arXiv preprint arXiv:2512.05349, 2025

    Zhenzhu Li and Aron Walsh. Platonic representation of foundation machine learning inter- atomic potentials.arXiv preprint arXiv:2512.05349, 2025

    arXiv 2025

  30. [30]

    Elign: Equivariant diffusion model alignment from foundational machine learning force fields.arXiv preprint arXiv:2601.21985, 2026

    Yunyang Li, Lin Huang, Luojia Xia, Wenhe Zhang, and Mark Gerstein. Elign: Equivariant diffusion model alignment from foundational machine learning force fields.arXiv preprint arXiv:2601.21985, 2026

    arXiv 2026

  31. [31]

    Unifying force pre- diction and molecular conformation generation through representation alignment

    Lucas Pinede, Soojung Yang, Juno Nam, and Rafael Gomez-Bombarelli. Unifying force pre- diction and molecular conformation generation through representation alignment. InICML 2025 Generative AI and Biology (GenBio) Workshop, 2025

    2025

  32. [32]

    Generative pseudo-force fields for molecular generation.arXiv preprint arXiv:2605.19050, 2026

    Stefaan Simon Pierre Hessmann, Khaled Kahouli, Stefan Gugler, Michael Plainer, Frank No ´e, Klaus-Robert M ¨uller, and Niklas Wolf Andreas Gebauer. Generative pseudo-force fields for molecular generation.arXiv preprint arXiv:2605.19050, 2026

    Pith/arXiv arXiv 2026

  33. [33]

    Score-based generative modeling through stochastic differential equations

    Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021

    2021

  34. [34]

    Learning multi- scale local conditional probability models of images

    Zahra Kadkhodaie, Florentin Guth, St ´ephane Mallat, and Eero P Simoncelli. Learning multi- scale local conditional probability models of images. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  35. [35]

    Visual autoregressive modeling: Scalable image generation via next-scale prediction

    Keyu Tian, Yi Jiang, Zehuan Yuan, BINGYUE PENG, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  36. [36]

    The optimal one dimensional periodic table: a modified pettifor chemical scale from data mining.New Journal of Physics, 18(9):093011, sep 2016

    Henning Glawe, Antonio Sanna, E K U Gross, and Miguel A L Marques. The optimal one dimensional periodic table: a modified pettifor chemical scale from data mining.New Journal of Physics, 18(9):093011, sep 2016

    2016

  37. [37]

    Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding.arXiv preprint arXiv:1807.03748, 2018. 14

    Pith/arXiv arXiv 2018

  38. [38]

    Jakob, Aron Walsh, Karsten Reuter, and Johannes T

    Konstantin S. Jakob, Aron Walsh, Karsten Reuter, and Johannes T. Margraf. Learning Crys- tallographic Disorder: Bridging Prediction and Experiment in Materials Discovery.Advanced Materials, 38(5):e14226, January 2026

    2026

  39. [39]

    Ilyes Batatia, David Peter Kovacs, Gregor N. C. Simm, Christoph Ortner, and Gabor Csanyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, Advances in Neural Information Processing Systems, 2022

    2022

  40. [40]

    Ilyes Batatia, Simon Batzner, D ´avid P ´eter Kov ´acs, Albert Musaelian, Gregor N. C. Simm, Ralf Drautz, Christoph Ortner, Boris Kozinsky, and G ´abor Cs´anyi. The design space of e(3)- equivariant atom-centered interatomic potentials.arXiv preprint arXiv:2205.06643, 2022

    Pith/arXiv arXiv 2022

  41. [41]

    Castelli, David D

    Ivano E. Castelli, David D. Landis, Kristian S. Thygesen, Søren Dahl, Ib Chorkendorff, Thomas F. Jaramillo, and Karsten W. Jacobsen. New cubic perovskites for one- and two- photon water splitting using the computational materials repository.Energy Environ. Sci., 5:9034–9043, 2012

    2012

  42. [42]

    Castelli, Thomas Olsen, Soumendu Datta, David D

    Ivano E. Castelli, Thomas Olsen, Soumendu Datta, David D. Landis, Søren Dahl, Kristian S. Thygesen, and Karsten W. Jacobsen. Computational screening of perovskite metal oxides for optimal solar light capture.Energy Environ. Sci., 5:5814–5819, 2012

    2012

  43. [43]

    Crystal- ite: A lightweight transformer for efficient crystal modeling.arXiv preprint arXiv:2604.02270, 2026

    Tin Had ˇzi Veljkovi´c, Joshua Rosenthal, Ivor Lonˇcari´c, and Jan-Willem van de Meent. Crystal- ite: A lightweight transformer for efficient crystal modeling.arXiv preprint arXiv:2604.02270, 2026

    Pith/arXiv arXiv 2026

  44. [44]

    Building normalizing flows with stochastic interpolants

    Michael Samuel Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  45. [45]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  46. [46]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and qiang liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  47. [47]

    Itai Gat, Tal Remez, Neta Shaul, Felix Kreuk, Ricky T. Q. Chen, Gabriel Synnaeve, Yossi Adi, and Yaron Lipman. Discrete flow matching. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  48. [48]

    Simple and ef- fective masked diffusion language models

    Subham Sekhar Sahoo, Marianne Arriola, Aaron Gokaslan, Edgar Mariano Marroquin, Alexander M Rush, Yair Schiff, Justin T Chiu, and V olodymyr Kuleshov. Simple and ef- fective masked diffusion language models. InThe Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  49. [49]

    E(n) equivariant graph neural networks

    V ´ıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E(n) equivariant graph neural networks. In Marina Meila and Tong Zhang, editors,Proceedings of the 38th International Conference on Machine Learning, volume 139 ofProceedings of Machine Learning Research, pages 9323–9332. PMLR, 18–24 Jul 2021

    2021

  50. [50]

    The mlip package: moment tensor potentials with mpi and active learning.Machine Learning: Science and Technology, 2(2):025002, dec 2020

    Ivan S Novikov, Konstantin Gubaev, Evgeny V Podryabinkin, and Alexander V Shapeev. The mlip package: moment tensor potentials with mpi and active learning.Machine Learning: Science and Technology, 2(2):025002, dec 2020

    2020

  51. [51]

    Courville

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron C. Courville. Film: Visual reasoning with a general conditioning layer. InAAAI, 2018

    2018

  52. [52]

    Anuroop Sriram, Benjamin Kurt Miller, Ricky T. Q. Chen, and Brandon M Wood. FlowLLM: Flow matching for material generation with large language models as base distributions. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  53. [53]

    C ´edric Villani.Optimal transport – Old and new, volume 338, pages xxii+973. 01 2008. 15

    2008

  54. [54]

    H. W. Kuhn. The hungarian method for the assignment problem.Naval Research Logistics Quarterly, 2(1-2):83–97, 1955

    1955

  55. [55]

    Alaya, Aur ´elie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L ´eo Gautheron, Nathalie T.H

    R ´emi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z. Alaya, Aur ´elie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L ´eo Gautheron, Nathalie T.H. Gayraud, Hicham Janati, Alain Rakotomamonjy, Ievgen Redko, An- toine Rolet, Antony Schutz, Vivien Seguy, Danica J. Sutherland, Romain Tavenard, Alexand...

    2021

  56. [56]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In C.J. Burges, L. Bottou, M. Welling, Z. Ghahramani, and K.Q. Weinberger, editors,Advances in Neural Information Processing Systems, volume 26. Curran Associates, Inc., 2013

    2013

  57. [57]

    Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample flow matching: Straightening flows with mini- batch couplings. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, editors,Proceedings of the 40th International Confer- ence on...

    2023

  58. [58]

    Expected sliced transport plans

    Xinran Liu, Rocio Diaz Martin, Yikun Bai, Ashkan Shahbazi, Matthew Thorpe, Akram Al- droubi, and Soheil Kolouri. Expected sliced transport plans. InThe Thirteenth International Conference on Learning Representations, 2025

    2025

  59. [59]

    Multiscale strategies for computing optimal transport

    Samuel Gerber and Mauro Maggioni. Multiscale strategies for computing optimal transport. Journal of Machine Learning Research, 18(72):1–32, 2017

    2017

  60. [60]

    Improved precision and recall metric for assessing generative models

    Tuomas Kynk ¨a¨anniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch´e-Buc, E. Fox, and R. Garnett, editors,Advances in Neural Informa- tion Processing Systems, volume 32. Curran Associates, Inc., 2019

    2019

  61. [61]

    Reliable fidelity and diversity metrics for generative models

    Muhammad Ferjad Naeem, Seong Joon Oh, Youngjung Uh, Yunjey Choi, and Jaejun Yoo. Reliable fidelity and diversity metrics for generative models. In Hal Daum´e III and Aarti Singh, editors,Proceedings of the 37th International Conference on Machine Learning, volume 119 ofProceedings of Machine Learning Research, pages 7176–7185. PMLR, 13–18 Jul 2020

    2020

  62. [62]

    Pac-bayesian contrastive unsupervised representation learning

    Kento Nozawa, Pascal Germain, and Benjamin Guedj. Pac-bayesian contrastive unsupervised representation learning. In Jonas Peters and David Sontag, editors,Proceedings of the 36th Conference on Uncertainty in Artificial Intelligence (UAI), volume 124 ofProceedings of Ma- chine Learning Research, pages 21–30. PMLR, 03–06 Aug 2020

    2020

  63. [63]

    Chris J. Pickard. Airss data for carbon at 10gpa and the c+n+h+o system at 1gpa, 2020

    2020

  64. [64]

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skinner, Gerbrand Ceder, and Kristin A. Persson. Commentary: The materials project: A materials genome approach to accelerating materials innovation.APL Materials, 1(1):011002, 07 2013

    2013

  65. [65]

    Ricky T. Q. Chen and Yaron Lipman. Flow matching on general geometries. InThe Twelfth International Conference on Learning Representations, 2024

    2024

  66. [66]

    Conditional wasserstein distances with applications in bayesian ot flow matching, 2025

    Jannis Chemseddine, Paul Hagemann, Gabriele Steidl, and Christian Wald. Conditional wasserstein distances with applications in bayesian ot flow matching, 2025. 16 A Primer on Optimal Transport We will briefly recall the basics of Optimal Transport (OT), following [11]; for a more theoretical, abstract overview, see [53]. Optimal Transport is concerned wit...

    2025

  67. [67]

    This is realized as a doubly stochastic matrix

    Step 1: Calculate the pairwise distance matrixC i,j =∥x i −g j∥and the OT plan π(µ, ν)∈R n,m betweenµ= 1 n Pn i=1 δxi andν= 1 m Pm i=1 δgi. This is realized as a doubly stochastic matrix

  68. [68]

    Step 2: Calculate the actual TNovD metric as TNovD(µ, ν) = nX i=1 mX j=1 πi,j(µ, ν)(ReLU(Ci,j −τ) +MReLU(τ−C i,j)). During step 2,τserves as a threshold that splitsC i,j into a memorization and a quality regime, which are supposed to reflect the above introduced novelty and quality coverage, respectively. Since 18 the memorization values are below the thr...

  69. [69]

    The training data set

  70. [70]

    The validation data set

  71. [71]

    Theβ-Goldilocks percentage, i.e., the relative number of points that are categorized into the Goldilocks zone. 21 From there on, we calculate the featurized versions of the training and the validation samples with respect to the two featurizersF id, Fmace, and find the optimal coupling with respect to the coupling C= 1 2 C id + 1 2 C mace. Then, we assign...