Pith. sign in

REVIEW 4 major objections 4 minor 63 references

A single distributional distance can separate memorization from low quality in generative models of crystals.

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 17:24 UTC pith:6IWNDECO

load-bearing objection TNovD is a genuinely useful distributional metric for material generative models, with convincing sanity checks; the main soft spot is a mathematically flawed heuristic for M, but the core idea holds up and deserves a real referee. the 4 major comments →

arxiv 2512.09514 v1 pith:6IWNDECO submitted 2025-12-10 cond-mat.mtrl-sci cs.LG

Transport Novelty Distance: A Distributional Metric for Evaluating Material Generative Models

Paul Hagemann , Simon M\"uller , Janine George , Philipp Benner This is my paper
classification cond-mat.mtrl-sci cs.LG MSC 49Q2268T07
keywords material generative modelsoptimal transportnovelty metricmemorization detectioncontrastive learninggraph neural networkscrystal structure generationdistributional evaluation
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.

The paper introduces Transport Novelty Distance (TNovD), a distributional metric that scores a material generative model in one number by jointly penalizing memorization and poor quality. It builds on optimal transport: generated crystal structures are embedded by a contrastively trained graph neural network, matched to training structures via an optimal coupling, and each matched distance is re-priced so that distances below a threshold count as memorization and distances above it count as low quality. The authors show on toy and corruption experiments that TNovD behaves as intended—it rises linearly as training data is leaked, increases with noise, shear, and atom substitutions, stays invariant to supercells, and gives sensible rankings of published generative models. The metric is deliberately not a training objective because a generator could game it; it is proposed as an evaluation tool.

Core claim

The paper's central claim is that a single distributional number, TNovD, can judge a material generative model on both quality and novelty at once. TNovD first couples the learned feature representations of training and generated crystal structures under ordinary Euclidean optimal transport, then re-prices each paired distance with a piecewise cost: distances below a threshold τ are penalized as memorization, distances above τ as poor quality. The threshold and exponent M are fit with a heuristic adapted from diffusion-model memorization literature, and the features come from an SE(3)-invariant graph neural network trained with a contrastive InfoNCE objective on the MP20 training set. The pa

What carries the argument

The central object is the piecewise cost function ω(x) = (τ−x)^M for x < τ and x−τ for x ≥ τ, applied to optimal-transport pair distances ||F(x_i) − F(g_{π*(i)})||. The coupling is computed once with standard Euclidean cost; the piecewise reweighting then separates a memorization regime (paired distances too close to training samples) from a quality regime (paired distances too far). The featurizer F is an SE(3)-invariant graph neural network trained with InfoNCE so that rotations, translations, and supercells of the same crystal map near each other while distinct materials are pushed apart; this learned embedding is the space in which closeness and distance are judged.

Load-bearing premise

The metric is only as good as the learned feature space: if Euclidean distance in that space does not mirror real crystal (dis)similarity, both the memorization cutoff and the quality scores misrank generators.

What would settle it

A decisive test: take a featurizer deliberately broken by training on shuffled atom labels, then rerun the paper's leaked-training experiment; if TNovD no longer rises with the fraction of copied training structures, the metric's memorization signal depends entirely on the embedding rather than on the distributional coupling. A cheaper check is to replace the 1/9 nearest-neighbour rule with human labels of memorized versus novel structures on a sample of generated structures and measure how often the τ split agrees.

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

If this is right

  • A memorizing generator can no longer hide behind Fréchet-type Gaussian summaries: reproduced training samples fall into the penalized memorization regime and inflate TNovD.
  • A generator that distorts structures through noise, shear, or substitutions is separated from a good generator by the quality regime, with larger penalties for more severe corruption.
  • Because supercell expansions map to the same features as the primitive cell, a model that only outputs supercell variants is not falsely penalized as novel.
  • Model rankings become decomposable into quality and memorization components: the paper finds ADiT ranks best, MatterGen is close to the validation set, and CDVAE ranks worst, with Chameleon showing the most memorization.
  • TNovD is explicitly not a training loss, since a generator could exploit the τ-sphere around training data to minimize it; it is intended for evaluation only.

Where Pith is reading between the lines

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

  • Because the construction is agnostic to the featurizer, the same two-regime optimal-transport cost should transfer to images and molecules given a contrastive embedding; a direct test would be to reproduce the memorization/noise U-shape on another domain.
  • The paper's observation that the quality component dominates for real models is ambiguous: either current materials generators rarely reproduce training data, or the 32-dimensional embedding is too coarse to catch near-duplicates; ablating embedding dimension and τ would separate these possibilities.
  • The τ threshold fitted by the 1/9 nearest-neighbour rule is global; making τ local or density-adaptive could improve detection of memorized outliers in dense regions of feature space.
  • Adding an energy or stability term, such as energy above hull, as a third penalty would let the same framework rank candidates by synthesizability, which the current metric explicitly does not claim to capture.

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

4 major / 4 minor

Summary. The paper introduces TNovD, a distributional score for evaluating material generative models that aims to penalize both memorization and low quality. The score is computed by (1) embedding training and generated crystal structures with an SE(3)-invariant GNN trained via InfoNCE contrastive learning, (2) solving an optimal-transport coupling in that feature space with Euclidean cost, and (3) applying a thresholded cost omega(d) = (tau - d)^M for d < tau (memorization regime) and d - tau for d >= tau (quality regime). The authors validate TNovD on a 2D flow-matching toy example, on MP20 validation data subjected to memorization injection, noise, shear, substitutions, and supercells, on the WBM substitution dataset, and on generated samples from six published models. They claim that a low TNovD indicates generated materials that are simultaneously novel and realistic.

Significance. If the method works as claimed, TNovD would be a valuable complement to existing per-instance SUN metrics and to Gaussian-assumption FID-type scores for materials. The paper ships code, includes carefully chosen sanity checks (memorization injection, deformations, substitutions), and the qualitative behavior in Fig. 5 is encouraging. However, the central hyperparameter calibration for M rests on an incorrect derivative, and the transfer of the memorization threshold from image diffusion to a learned crystal feature space is not directly validated. These issues affect the interpretation of the metric, not merely its presentation.

major comments (4)
  1. [Heuristic for the Hyperparameter] The derivative used to calibrate M is incorrect. With omega(d) = (tau - d)^M for d < tau and d - tau for d >= tau, the derivative of TNovD = (1/n) sum_i omega(d_i) with respect to tau is M E[(tau - D)^{M-1} 1_{D <= tau}] - P(D > tau), not M P(D <= tau) - P(D > tau). Consequently, the equilibrium condition partial_tau TNovD = 0 does not lead to M = P(D > tau)/P(D <= tau). Since M controls the relative weight of memorization and quality components, this is load-bearing for the metric's interpretation. The calibration of M is unprincipled as written; the authors should correct the derivation, recalibrate M, and report the resulting values, or adopt an explicit alternative objective for M.
  2. [Heuristic for the Hyperparameter / Numerical Results] The threshold tau is calibrated by transplanting the image-diffusion memorization rule d1 <= d2/9 into the GNN feature space and optimizing a 1-NN classifier on MP20 train. This transfer is not validated: the GNN embedding has arbitrary scale, is trained only on MP20 train, and there is no direct evidence that feature-space distances at the learned tau separate memorized from novel materials. Fig. 5(a) shows that TNovD responds monotonically to injected duplicates, but sensitivity is not classification accuracy. A direct validation, e.g. known duplicate/near-duplicate pairs or retrieval-precision experiments, is needed before TNovD can be interpreted as a joint novelty/quality score.
  3. [Heuristic for the Hyperparameter / Main Benchmarks] The fitted numerical values of tau and M are not reported anywhere, and no sensitivity analysis is provided. The U-shaped behavior in Fig. 3(f) and the monotone responses in Fig. 5 are presented as validation of the metric, but without these values or a sweep over plausible (tau, M) one cannot tell whether the rankings in Fig. 7 are robust or driven by a particular calibration. Please report tau, M, and the dependence of the main benchmark conclusions on them.
  4. [Abstract and Conclusion] The abstract and introduction state that TNovD judges 'quality' and 'novelty' and can identify 'realistic' materials. However, TNovD only measures distributional similarity to training-set embeddings; it does not directly measure thermodynamic stability or synthesizability, as the authors themselves concede ('it is not clear that our featurizer accurately captures stability properties'). This is an acknowledged scope limitation, but the language in the abstract and contribution should be softened accordingly: a low TNovD indicates feature-space novelty and fidelity to the training distribution, not physical realizability.
minor comments (4)
  1. [Throughout] The paper alternates between calling TNovD a 'distance' and noting it is 'not a distance in the mathematical sense.' Consider using 'score' consistently to avoid confusion.
  2. [Numerical Results] Typographical issues: 'Pertubations' should be 'Perturbations'; 'Chemeleon' should be 'Chameleon'; 'Assumably' should be 'Presumably'; 'CDV AE' should be 'CDVAE'.
  3. [Behavior on the WBM-Dataset] The sentence 'all stable structures with energy above hull values below 0 were extracted' is unclear; energy above hull is normally non-negative, so please clarify the selection criterion.
  4. [Transport Novelty Distance for common generative models] Fig. 7 reports a single benchmark run per model without error bars or multiple seeds. Since the rankings are a central result, a statement about stochastic variability would strengthen the claims.

Circularity Check

2 steps flagged

Minor self-definitional sanity checks; central TNovD derivation is not circular.

specific steps
  1. self definitional [Section 'The Transport Novelty Distance', definition of TNovD/ω; Section 'Numerical Results', Fig. 5(a)]
    "TNovD(PD,PG)= 1/n Σ_i ω(∥F(x_i)−F(g_π*(i))∥) where ω(x)= (τ−x)^M, x<τ (memorization regime), (x−τ), x≥τ (quality regime). ... In the Fig. 5 (a), we observe that our TNovD performs excellently in recognizing memorized data, with an almost perfectly linear increase for increasing amounts of leaked training data."

    By the displayed definition, any pair with feature distance d<τ contributes (τ−d)^M, so an exact training replicate (d=0) contributes τ^M>0. Replacing validation samples with training samples (or rotations/translations thereof, which the InfoNCE embedding maps to the same features) therefore forces the average to rise approximately linearly with the replacement fraction. The 'recognition of memorized data' in Fig. 5(a) is an algebraic consequence of the cost function's memorization regime, not an independent empirical confirmation that the metric captures memorization.

  2. self definitional [Section 'A GNN featurizer' / 'Contrastive Loss' and Section 'Numerical Results', Fig. 5(d)]
    "To account for this phenomenon, we added supercells as augmentations to our positive examples in the InfoNCE loss. The GNN is therefore trained to map the supercell augments to the same features as the original cell. ... Further, we also see that our model is perfectly invariant to supercells (Fig. 5 (d))."

    Supercells were included as positive pairs in the InfoNCE objective, forcing F(supercell)≈F(parent). TNovD is computed only from Euclidean distances between these embeddings, so distances involving a supercell coincide with those of the parent cell. Reporting 'perfect invariance' in Fig. 5(d) is reading the training objective back out of the metric; it does not validate the metric against an independent notion of supercell equivalence.

full rationale

The paper does not derive TNovD from its inputs in a circular way: TNovD is explicitly constructed as a piecewise cost on optimal-transport distances, and the model-ranking results in Fig. 7 depend on the contrastively trained GNN embedding, which is trained independently of the metric evaluation. I found no load-bearing self-citation chain and no imported uniqueness theorem. The main caveats are correctness risks rather than circularity: the M heuristic's derivative is miscomputed, and the 1/9 memorization rule is transferred from image diffusion without validation in crystal feature space. The two sanity checks flagged above are self-definitional, but they concern peripheral validations, not the central derivation; hence a modest score.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The metric rests on free hyperparameters (tau, M) fitted to the training distribution, plus a learned embedding whose metric properties are assumed rather than validated. No new physical entities are introduced. The most fragile items are the transferred memorization rule and the approximate M derivation.

free parameters (5)
  • tau (memorization threshold) = not reported (grid-search minimum on MP20 train)
    Chosen by grid search to approximate the 1-NN memorization rule d1 <= d2/9 from diffusion literature; controls the boundary between memorization and quality regimes.
  • M (memorization penalty exponent) = M = P(X_dist > tau) / P(X_dist <= tau), estimated on a split of MP20 train
    Derived from an approximate zero-derivative equilibrium condition; the derivation drops the (tau-x)^(M-1) factor, so the value is approximate.
  • GNN feature dimension = 32
    Chosen by hand; no ablation is reported.
  • InfoNCE temperature T = 0.1
    Standard choice; no ablation is reported.
  • training epochs for GNN featurizer = 10
    Chosen by hand; no ablation is reported.
axioms (5)
  • domain assumption Memorization rule ||x - x(1)||^2 <= (1/9) ||x - x(2)||^2 defines memorization in feature space
    Imported from diffusion-memorization papers [50-52]; applied without empirical validation to crystal feature distances.
  • domain assumption The InfoNCE-trained GNN embedding produces Euclidean distances that reflect material dissimilarity
    Central modeling assumption of the featurizer; the paper provides no independent validation of the metric properties of the embedding.
  • ad hoc to paper For two 'good' distributions, tau is optimal and the derivative of TNovD w.r.t. tau is zero
    Assumed in the heuristic to derive M; the derivative formula is not exact because it ignores the (tau-x)^(M-1) factor.
  • domain assumption MP20 validation structures are all novel relative to MP20 train
    The paper assumes no duplicate or near-duplicate structures across the split; used to interpret the validation set as a good-generator proxy.
  • standard math Kantorovich optimal-transport couplings exist and can be computed discretely
    Standard OT theory used to define and compute TNovD; the paper relies on POT for implementation.

pith-pipeline@v1.3.0-alltime-deepseek · 16096 in / 13769 out tokens · 136053 ms · 2026-08-03T17:24:19.785823+00:00 · methodology

0 comments
read the original abstract

Recent advances in generative machine learning have opened new possibilities for the discovery and design of novel materials. However, as these models become more sophisticated, the need for rigorous and meaningful evaluation metrics has grown. Existing evaluation approaches often fail to capture both the quality and novelty of generated structures, limiting our ability to assess true generative performance. In this paper, we introduce the Transport Novelty Distance (TNovD) to judge generative models used for materials discovery jointly by the quality and novelty of the generated materials. Based on ideas from Optimal Transport theory, TNovD uses a coupling between the features of the training and generated sets, which is refined into a quality and memorization regime by a threshold. The features are generated from crystal structures using a graph neural network that is trained to distinguish between materials, their augmented counterparts, and differently sized supercells using contrastive learning. We evaluate our proposed metric on typical toy experiments relevant for crystal structure prediction, including memorization, noise injection and lattice deformations. Additionally, we validate the TNovD on the MP20 validation set and the WBM substitution dataset, demonstrating that it is capable of detecting both memorized and low-quality material data. We also benchmark the performance of several popular material generative models. While introduced for materials, our TNovD framework is domain-agnostic and can be adapted for other areas, such as images and molecules.

Figures

Figures reproduced from arXiv: 2512.09514 by Janine George, Paul Hagemann, Philipp Benner, Simon M\"uller.

Figure 1
Figure 1. Figure 1: Graphical explanation of the complete workflow used to derive the Transport Novelty Distance [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visualization of an OT coupling be￾tween two discrete distributions. The connect￾ing lines indicate the coupling, i.e., the sample assignments. One might ask whether OT is limited to datasets of the same size. Luckily, the framework is strictly more general and also works for continuous data or datasets of different sizes. This is important for our application, where the total number of materials in the tr… view at source ↗
Figure 3
Figure 3. Figure 3: (a): True uniform distribution on the circle. (b) to (d): Generated samples for increasing smoothing [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of optimal-transport matchings between training data [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Different toy experiments performed to test the TNovD behavior for specific test cases. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Transport Novelty Distance for the full MP20 data (train + validation + test set) and all five steps of [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: TNovD for the MP20 validation set, as well as for 10,000 predicted samples from six different [PITH_FULL_IMAGE:figures/full_fig_p011_7.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

63 extracted references · 6 canonical work pages

  1. [1]

    Houssam Metni, Laura Ruple, Lauren N. Walters, Luca Torresi, Jonas Teufel, Henrik Schopmans, Jona ¨Ostreicher, Yumeng Zhang, Marlen Neubert, Yuri Koide, Kevin Steiner, Paul Link, Lukas B¨ar, Mariana Petrova, Gerbrand Ceder, and Pascal Friederich. Generative models for crystalline materials, 2025. URLhttps://arxiv.org/abs/2511.22652

    arXiv 2025

  2. [2]

    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 generative modelling of molecules and materials. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceed...

    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. URLhttps://openreview.net/forum?id=03RLpj-tc_

    2022

  4. [4]

    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 T...

    doi:10.1038/s41586-025-08628-5 2025

  5. [5]

    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. URLhttps://openreview.net/forum?id=DNdN26m2Jk

    2023

  6. [6]

    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 Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In I. Guyon, U. Von 12 Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems, volume 30. Curran Associa...

    2017

  7. [7]

    Fr´echet chemnet distance: A metric for generative models for molecules in drug discovery.Journal of Chemical Information and Modeling, 58(9):1736–1741, 2018

    Kristina Preuer, Philipp Renz, Thomas Unterthiner, Sepp Hochreiter, and G¨ unter Klambauer. Fr´echet chemnet distance: A metric for generative models for molecules in drug discovery.Journal of Chemical Information and Modeling, 58(9):1736–1741, 2018. doi: 10.1021/acs.jcim.8b00234. URL https://doi.org/10.1021/acs.jcim.8b00234. PMID: 30118593

    doi:10.1021/acs.jcim.8b00234 2018

  8. [8]

    Vector field oriented diffusion model for crystal material generation.Proceedings of the AAAI Conference on Artificial Intelligence, 38(20): 22193–22201, Mar

    Astrid Klipfel, Ya ¨el Fregier, Adlane Sayede, and Zied Bouraoui. Vector field oriented diffusion model for crystal material generation.Proceedings of the AAAI Conference on Artificial Intelligence, 38(20): 22193–22201, Mar. 2024. doi: 10.1609/aaai.v38i20.30224. URL https://ojs.aaai.org/index.php/AAAI/article/view/30224

    work page doi:10.1609/aaai.v38i20.30224 2024

  9. [9]

    Jakob, Aron Walsh, Karsten Reuter, and Johannes T

    Konstantin S. Jakob, Aron Walsh, Karsten Reuter, and Johannes T. Margraf. Learning crystallographic disorder: Bridging prediction and experiment in materials discovery.Advanced Materials, n/a(n/a): e14226. doi: https://doi.org/10.1002/adma.202514226. URL https://advanced.onlinelibrary.wiley.com/doi/abs/10.1002/adma.202514226

    work page doi:10.1002/adma.202514226

  10. [10]

    Continued challenges in high-throughput materials predictions: Mattergen predicts compounds from the training dataset, 2025

    Mikkel Juelsholt. Continued challenges in high-throughput materials predictions: Mattergen predicts compounds from the training dataset, 2025. URL https://chemrxiv.org/engage/chemrxiv/article-details/685057d9c1cb1ecda05e9aec

    2025

  11. [11]

    Muhammad Usman Akbar, Wuhao Wang, and Anders Eklund. Beware of diffusion models for synthesizing medical images—a comparison with gans in terms of memorizing brain mri and chest x-ray images.Machine Learning: Science and Technology, 6(1):015022, jan 2025. doi: 10.1088/2632-2153/ad9a3a. URLhttps://doi.org/10.1088/2632-2153/ad9a3a

    work page doi:10.1088/2632-2153/ad9a3a 2025

  12. [12]

    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. InInternational Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=PxTIG12RRHS

    2021

  13. [13]

    C ´edric Villani.Optimal transport – Old and new, volume 338, pages xxii+973. 01 2008. doi: 10.1007/978-3-540-71050-9

    doi:10.1007/978-3-540-71050-9 2008

  14. [14]

    De Gruyter, Berlin, Boston, 2017

    Simone Di Marino, Augusto Gerolin, and Luca Nenna.Optimal transportation theory with repulsive costs, pages 204–256. De Gruyter, Berlin, Boston, 2017. ISBN 9783110430417. doi: 10.1515/9783110430417-010. URLhttps://doi.org/10.1515/9783110430417-010

    doi:10.1515/9783110430417-010 2017

  15. [15]

    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. URLhttps://proceedings.mlr.press/v139/satorras21a.html

    2021

  16. [16]

    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

    Pith/arXiv arXiv 2018

  17. [17]

    Mastej, and Aron Walsh

    Masahiro Negishi, Hyunsoo Park, Kinga O. Mastej, and Aron Walsh. Continuous uniqueness and novelty metrics for generative modeling of inorganic crystals.arXiv preprint arXiv:2510.12405, 2025

    arXiv 2025

  18. [18]

    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. ISSN 2166-532X. doi: 10.1063/1.481232...

    doi:10.1063/1.4812323 2013

  19. [19]

    Hai-Chen Wang, Miguel A. L. Marques, and Silvana Botti. Predicting stable crystalline compounds using chemical similarity.npj Comput Mater, 7, 2021. doi: https://doi.org/10.1038/s41524-020-00481-6

    doi:10.1038/s41524-020-00481-6 2021

  20. [20]

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

    Alexandre Duval, Siddharth Betala, Samuel P. Gleason, Andy Xu, Georgia Channing, Daniel Levy, Ali Ramlaoui, Cl´ementine Fourrier, Chaitanya K. Joshi, Nikita Kazeev, S´ekou-Oumar Kaba, F´elix Therrien, Alex Hern´andez-Garc´ıa, Roc´ıo Mercado, and N M Anoop Krishnan. Lemat-genbench: Bridging the gap between crystal generation and materials discovery. InAI f...

    2025

  21. [21]

    Sterling G Baird, Hasan M Sayeed, Joseph Montoya, and Taylor D Sparks. matbench-genmetrics: A python library for benchmarking crystal structure generative models using time-based splits of materials project structures.Journal of Open Source Software, 9(97):5618, 2024

    2024

  22. [22]

    Szymanski and Christopher J

    Nathan J. Szymanski and Christopher J. Bartel. Establishing baselines for generative discovery of inorganic crystals.arXiv preprint arXiv:2501.02144, 2025

    Pith/arXiv arXiv 2025

  23. [23]

    Hargreaves, Matthew S

    Cameron J. Hargreaves, Matthew S. Dyer, Michael W. Gaultois, Vitaliy A. Kurlin, and Matthew J. Rosseinsky. The earth mover’s distance as a metric for the space of inorganic compositions.Chemistry of Materials, 32(24):10610–10620, 2020. doi: 10.1021/acs.chemmater.0c03381. URL https://doi.org/10.1021/acs.chemmater.0c03381

    doi:10.1021/acs.chemmater.0c03381 2020

  24. [24]

    Hegde, Kevin V

    Anthony Onwuli, Ashish V. Hegde, Kevin V. T. Nguyen, Keith T. Butler, and Aron Walsh. Element similarity in high-dimensional materials representations.Digital Discovery, 2:1558–1564, 2023. doi: 10.1039/D3DD00121K. URLhttp://dx.doi.org/10.1039/D3DD00121K

    work page doi:10.1039/d3dd00121k 2023

  25. [25]

    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 symmetry. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors,Proceedings of the 42nd International ...

  26. [26]

    Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models

    George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances i...

    2023

  27. [27]

    Understanding and mitigating memorization in generative models via sharpness of probability landscapes

    Dongjae Jeon, Dueun Kim, and Albert No. Understanding and mitigating memorization in generative models via sharpness of probability landscapes. InForty-second International Conference on Machine Learning, 2025. URLhttps://openreview.net/forum?id=EW2JR5aVLm

    2025

  28. [28]

    Score-based generative models detect manifolds

    Jakiw Pidstrigach. Score-based generative models detect manifolds. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors,Advances in Neural Information Processing Systems,

  29. [29]

    An interpretable evaluation of entropy-based novelty of generative models.arXiv preprint arXiv:2402.17287, 2024

    Jingwei Zhang, Cheuk Ting Li, and Farzan Farnia. An interpretable evaluation of entropy-based novelty of generative models.arXiv preprint arXiv:2402.17287, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Feature likelihood score: Evaluating the generalization of generative models using samples

    Marco Jiralerspong, Joey Bose, Ian Gemp, Chongli Qin, Yoram Bachrach, and Gauthier Gidel. Feature likelihood score: Evaluating the generalization of generative models using samples. InThirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=l2VKZkolT7

    2023

  31. [31]

    Density estimation using real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using real NVP. In International Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=HkpbnH9lx

    2017

  32. [32]

    Computational optimal transport: With applications to data science

    Gabriel Peyr ´e and Marco Cuturi. Computational optimal transport: With applications to data science. Foundations and Trends®in Machine Learning, 11(5-6):355–607, 2019. ISSN 1935-8237. doi: 10.1561/2200000073

    doi:10.1561/2200000073 2019

  33. [33]

    Improving and generalizing flow-based generative models with minibatch optimal transport.Transactions on Machine Learning Research, 2024

    Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport.Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URLhttps://openreview.net/forum?id=CD9Snc73AW. Expert Certification

    2024

  34. [34]

    OT-Flow: Fast and accurate continuous normalizing flows via optimal transport

    Derek Onken, Samy Wu Fung, Xingjian Li, and Lars Ruthotto. OT-Flow: Fast and accurate continuous normalizing flows via optimal transport. InAAAI Conference on Artificial Intelligence, volume 35, pages 9223–9232, May 2021. URL https://ojs.aaai.org/index.php/AAAI/article/view/17113. 14

    2021

  35. [35]

    The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009

    Franco Scarselli, Marco Gori, Ah Chung Tsoi, Markus Hagenbuchner, and Gabriele Monfardini. The graph neural network model.IEEE Transactions on Neural Networks, 20(1):61–80, 2009. doi: 10.1109/TNN.2008.2005605

    arXiv 2009

  36. [36]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. InInternational Conference on Learning Representations, 2017. URL https://openreview.net/forum?id=SJU4ayYgl

    2017

  37. [37]

    Wiltschko

    Benjamin Sanchez-Lengeling, Emily Reif, Adam Pearce, and Alexander B. Wiltschko. A gentle introduction to graph neural networks.Distill, 2021. doi: 10.23915/distill.00033. https://distill.pub/2021/gnn-intro

    doi:10.23915/distill.00033 2021

  38. [38]

    Schoenholz, Patrick F

    Justin Gilmer, Samuel S. Schoenholz, Patrick F. Riley, Oriol Vinyals, and George E. Dahl. Neural message passing for quantum chemistry. In Doina Precup and Yee Whye Teh, editors,Proceedings of the 34th International Conference on Machine Learning, volume 70 ofProceedings of Machine Learning Research, pages 1263–1272. PMLR, 06–11 Aug 2017. URL https://proc...

    2017

  39. [39]

    Frey, and Lin Li

    Teddy Koker, Keegan Quigley, Will Spaeth, Nathan C. Frey, and Lin Li. Graph contrastive learning for materials.arXiv preprint arXiv:2211.13408, 2022

    Pith/arXiv arXiv 2022

  40. [40]

    Csi: Novelty detection via contrastive learning on distributionally shifted instances

    Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. InAdvances in Neural Information Processing Systems, 2020

    2020

  41. [41]

    S. R. Hall, F. H. Allen, and I. D. Brown. The crystallographic information file (cif): a new standard archive file for crystallography.Acta Crystallographica Section A, 47(6):655–685, 1991. doi: https://doi.org/10.1107/S010876739101067X. URL https://onlinelibrary.wiley.com/doi/abs/10.1107/S010876739101067X

    doi:10.1107/s010876739101067x 1991

  42. [42]

    Alaya, Aur˜A©lie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L˜A©o Gautheron, Nathalie T.H

    R ´emi Flamary, Nicolas Courty, Alexandre Gramfort, Mokhtar Z. Alaya, Aur˜A©lie Boisbunon, Stanislas Chambon, Laetitia Chapel, Adrien Corenflos, Kilian Fatras, Nemo Fournier, L˜A©o Gautheron, Nathalie T.H. Gayraud, Hicham Janati, Alain Rakotomamonjy, Ievgen Redko, Antoine Rolet, Antony Schutz, Vivien Seguy, Danica J. Sutherland, Romain Tavenard, Alexander...

  43. [43]

    Chevrier, Kristin A

    Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L. Chevrier, Kristin A. Persson, and Gerbrand Ceder. Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis. Computational Materials Science, 68:314–319, February 2013. ISSN 09270256. URL...

    2013

  44. [44]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-perfo...

    2019

  45. [45]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. InICLR Workshop on Representation Learning on Graphs and Manifolds, 2019

    2019

  46. [46]

    Lenssen, and Jure Leskovec

    Matthias Fey, Jinu Sunil, Akihiro Nitta, Rishi Puri, Manan Shah, Blaˇ z Stojanoviˇc, Ramona Bendias, Alexandria Barghi, Vid Kocijan, Zecheng Zhang, Xinwei He, Jan E. Lenssen, and Jure Leskovec. PyG 2.0: Scalable learning on real world graphs. InTemporal Graph Learning Workshop @ KDD, 2025

    2025

  47. [47]

    Closed-form diffusion models

    Christopher Scarvelis, Haitz S ´aez de Oc´ariz Borde, and Justin Solomon. Closed-form diffusion models. Transactions on Machine Learning Research, 2025. ISSN 2835-8856. URL https://openreview.net/forum?id=JkMifr17wc

    2025

  48. [48]

    Diffusion models and the manifold hypothesis: Log-domain smoothing is geometry adaptive.arXiv preprint arXiv:2510.02305, 2025

    Tyler Farghly, Peter Potaptchik, Samuel Howard, George Deligiannidis, and Jakiw Pidstrigach. Diffusion models and the manifold hypothesis: Log-domain smoothing is geometry adaptive.arXiv preprint arXiv:2510.02305, 2025. 15

    arXiv 2025

  49. [49]

    Colocalization for super-resolution microscopy via optimal transport.Nature computational science, 1(3):199–211, 2021

    Carla Tameling, Stefan Stoldt, Till Stephan, Julia Naas, Stefan Jakobs, and Axel Munk. Colocalization for super-resolution microscopy via optimal transport.Nature computational science, 1(3):199–211, 2021

    2021

  50. [50]

    On the edge of memorization in diffusion models.arXiv preprint arXiv:2508.17689, 2025

    Sam Buchanan, Druv Pai, Yi Ma, and Valentin De Bortoli. On the edge of memorization in diffusion models.arXiv preprint arXiv:2508.17689, 2025

    Pith/arXiv arXiv 2025

  51. [51]

    Provable separations between memorization and generalization in diffusion models.arXiv preprint arXiv:2511.03202, 2025

    Zeqi Ye, Qijie Zhu, Molei Tao, and Minshuo Chen. Provable separations between memorization and generalization in diffusion models.arXiv preprint arXiv:2511.03202, 2025

    arXiv 2025

  52. [52]

    TaeHo Yoon, Joo Young Choi, Sehyun Kwon, and Ernest K. Ryu. Diffusion probabilistic models generalize when they fail to memorize. InICML 2023 Workshop on Structured Probabilistic Inference & Generative Modeling, 2023. URLhttps://openreview.net/forum?id=shciCbSk9h

    2023

  53. [53]

    Schnet: A continuous-filter convolutional neural network for modeling quantum interactions

    Kristof Sch¨ utt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Stefan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M¨ uller. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Informati...

    2017

  54. [54]

    Grossman

    Tian Xie and Jeffrey C. Grossman. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties.Phys. Rev. Lett., 120:145301, Apr 2018. doi: 10.1103/PhysRevLett.120.145301. URL https://link.aps.org/doi/10.1103/PhysRevLett.120.145301

    doi:10.1103/physrevlett.120.145301 2018

  55. [55]

    Cheetham and Ram Seshadri

    Anthony K. Cheetham and Ram Seshadri. Artificial intelligence driving materials discovery? perspective on the article: Scaling deep learning for materials discovery.Chemistry of Materials, 36(8): 3490–3495, 2024. doi: 10.1021/acs.chemmater.4c00643

    doi:10.1021/acs.chemmater.4c00643 2024

  56. [56]

    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. URL https://openreview.net/forum?id=jkvZ7v4OmP

    2024

  57. [57]

    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. doi: https://doi.org/10.1038/s41467-025-59636-y

    doi:10.1038/s41467-025-59636-y 2025

  58. [58]

    H. Park, K. O. Mastej, P. Detrattanawichai, R. Nduma, and A. Walsh. Closing the synthesis gap in computational materials design.ChemRxiv, 2025. doi: 10.26434/chemrxiv-2025-sbc0c-v2. URL https://doi.org/10.26434/chemrxiv-2025-sbc0c-v2

    work page doi:10.26434/chemrxiv-2025-sbc0c-v2 2025

  59. [59]

    Syncotrain: a dual classifier pu-learning framework for synthesizability prediction.Digital Discovery, 4:1437–1448, 2025

    Sasan Amariamir, Janine George, and Philipp Benner. Syncotrain: a dual classifier pu-learning framework for synthesizability prediction.Digital Discovery, 4:1437–1448, 2025. doi: 10.1039/D4DD00394B. URLhttp://dx.doi.org/10.1039/D4DD00394B

    work page doi:10.1039/d4dd00394b 2025

  60. [60]

    A foundation model for atomistic materials chemistry.The Journal of chemical physics, 163(18), 2025

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M Elena, D ´avid P Kov´acs, Janosh Riebesell, Xavier R Advincula, Mark Asta, Matthew Avaylon, William J Baldwin, et al. A foundation model for atomistic materials chemistry.The Journal of chemical physics, 163(18), 2025. 16

    2025

  61. [2021]

    URLhttp://jmlr.org/papers/v22/20-451.html

  62. [2022]

    URLhttps://openreview.net/forum?id=AiNrnIrDfD9

  63. [2025]

    URLhttps://proceedings.mlr.press/v267/ekstrom-kelvinius25a.html