Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Self-Supervised Learning for Ordered Three-Dimensional Structures

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read Self-supervised geometric models can distinguish two pre-crystallization liquids that classical order parameters cannot.

desk verdict A promising SSL framework for particle-centered structures whose headline LA/LB separation rests on a single un-replicated snapshot pair and needs error bars before it is trusted. read the letter →

arxiv 2411.14680 v1 pith:5JFDCMYV submitted 2024-11-22 cs.LG

Matthew Spellings , Maya Martirossyan , Julia Dshemuchadse This is my paper · ORCID
classification cs.LG
keywords self-supervisedlearninggeometricalgebraattentionnetworksrotationequivariance3Dstructureembeddingsorderparametersself-assemblytransfercrystallographicprototypes
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether self-supervised learning on ideal crystal structures can produce representations of three-dimensional local order that transfer to messy, disordered systems. It proposes six geometric pretext tasks—frame classification, denoising, autoencoding, shift identification, noisy bond classification, and nearest bond regression—and trains rotation- and permutation-equivariant networks on Gaussian-perturbed crystal prototype unit cells with no human labels. The central empirical result is that the frame-classification model, trained only on ideal crystals, embeds two pre-crystallization liquids (LA/LB) with ROC AUC 0.962, while Steinhardt Q and spherical-harmonic baselines score about 0.506 and 0.505. A sympathetic reader would care because distinguishing liquids that will form different solids is a nearly open problem in materials physics, and transfer from idealized data to simulations is evidence of generic learned order features.

What carries the argument

The architecture is a Geometric Algebra Attention (GAlA) layer: each input bond is represented as a vector-valued multivector paired with a type embedding, pairwise geometric products $p_{ij}$ are reduced to rotation-invariant attributes such as vector lengths, dot products, and bivector magnitudes, and attention weights combine those attributes with node embeddings to produce both rotation-equivariant and rotation-invariant outputs. Stacking three such layers forms a core that is equivariant under rotations and permutations of the input point cloud, which is what allows embeddings trained on one structure family to transfer to another. The pretext tasks force the core to keep geometric information: frame classification labels ideal structure prototypes after a permutation-invariant reduction, denoising asks the network to remove Gaussian perturbations, shift identification predicts a rigid displacement, and the remaining tasks similarly constrain the learned representation.

What would settle it

A reader could settle whether the LA/LB result is robust by rerunning the zero-shot evaluation on several independent cooling trajectories (different random seeds and potential parameters) for the cF4-Cu and hP2-Mg systems and checking whether the frame-classification AUC stays near 0.962 while the Steinhardt Q baseline stays near 0.506; a drop toward chance would falsify the transfer claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that self-supervised tasks defined on perturbed ideal unit cells teach equivariant networks a transferable notion of local crystalline order, one that is more informative for subtle disordered distinctions than conventional bond-orientational order parameters. Specifically, evaluating embeddings as zero-shot order parameters on a binary separation between liquids that will later crystallize into the cF4-Cu and hP2-Mg structures gives AUC 0.962 for frame classification and 0.859 for autoencoding, versus 0.506 for Steinhardt Q and 0.505 for local spherical harmonics. The frame-classification model was trained only on 50 one-component crystal prototypes with added Gaussian noise, then applied without retraining to molecular dynamics trajectories of self-assembling systems.

Load-bearing premise

The load-bearing premise is that each structure pair is represented by one cooling trajectory and that the 20-nearest-neighbor local environments in those simulations capture the same distinguishing signal that a model pretrained on ideal crystals has learned; if the trajectory sample is not representative, the LA/LB gap may not reproduce.

Editorial extensions

If this is right

  • If the central claim is right, self-supervised models trained on ideal crystals can act as zero-shot order parameters for simulations, identifying phase changes in trajectories without any labeled data from the simulated system.
  • The LA/LB result means learned geometric embeddings can resolve differences in local order that Steinhardt Q and spherical-harmonic descriptors average away, opening a route to studying competing liquid motifs before crystallization.
  • The transfer results indicate that pretraining on geometric pretext tasks reduces the amount of labeled data needed to classify frames or structures, with most source tasks beating direct training when all weights are fine-tuned.
  • Task-specific cores are complementary: localized tasks such as nearest bond regression and shift identification learn different information than autoencoding and denoising, so the full set of tasks constitutes a pretraining toolkit rather than any single one.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: nothing in the paper rules out that the LA/LB signal is tied to the particular final crystal pair cF4-Cu versus hP2-Mg; a natural stress test is to run the same zero-shot evaluation on several competing-liquid pairs with varied motif similarity.
  • Editorial inference: because the architecture retains type embeddings in its input, the same pretext tasks could be run on multicomponent mixtures to test whether learned order features survive chemical heterogeneity, which the current single-component results do not address.
  • Editorial inference: the paper mentions but does not implement concatenating the cores of all six tasks; if the learned representations are complementary, such a combined embedding is a concrete next step that could improve the weakest transfer rows in Figure 4.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces six self-supervised geometric tasks for particle-centered three-dimensional point clouds, together with a rotation- and permutation-equivariant network built from geometric algebra attention layers. The tasks are trained on both idealized AFLOW crystal prototypes and molecular dynamics trajectories, and the resulting embeddings are evaluated as order parameters for distinguishing gases, liquids, and solids. The paper also reports transfer learning among all pairs of tasks. The central empirical claim is that a frame-classification model trained only on Gaussian-perturbed ideal unit cells yields embeddings that distinguish two pre-crystallization liquids (LA/LB) with ROC AUC 0.962, compared with 0.506 for Steinhardt Q and 0.505 for local spherical harmonics.

Significance. If the headline LA/LB result is robust, the paper is a significant contribution to structural analysis of disordered assemblies: it would show that self-supervised equivariant representations can capture liquid preordering that classical bond-orientational order parameters miss, and that such representations transfer across structural tasks. The paper has concrete strengths: it proposes a coherent family of tasks, releases code and trajectories, compares directly against standard materials-science baselines, and is honest about some convergence difficulties in transfer learning. The evaluation is not circular, because the SSL training labels are defined independently of the target LA/LB comparison. However, the central quantitative claim currently rests on a single snapshot pair with no uncertainty quantification, and an appendix statement about the endpoint solids raises a question about what signal the model can possibly use. These issues are fixable, but they are load-bearing for the paper's main scientific claim.

major comments (3)
  1. [Results, Table 1] The headline LA/LB AUC (0.962 vs 0.506/0.505) is computed from one pair of snapshots, each from a single cooling trajectory, with no error bars, no replicate trajectories, and no test of statistical significance. Because the score is the first principal component of the embedding, a single dominant axis can produce a large AUC on one snapshot pair even if the representation does not capture a stable structural distinction. Please report AUCs with confidence intervals (bootstrap over particles and/or independent trajectories), evaluate multiple snapshots along each cooling curve, and include a significance test for the AUC difference against the baselines.
  2. [Appendix F, item 5, with Table 1] The paper states that cF4-Cu and hP2-Mg are 'very similar, having identical representations when using nearest-neighbor graphs.' Since the frame-classification model is trained on 20-nearest-neighbor point clouds of exactly these ideal prototypes (Appendix B), it is unclear what input signal allows it to separate the two liquids that crystallize into these solids. If the ideal solids have identical inputs, the learned representation cannot contain a graph-level distinction between them; the LA/LB separation would have to come from subtle geometric details of the liquid environments. Please clarify what 'identical representations' means for the model input, and demonstrate that the separation is reproducible across liquid snapshots and trajectories rather than an artifact of one snapshot.
  3. [Appendix E, Tables 4–8] Many transfer-learning values are orders of magnitude larger than any sensible error for the target task (e.g., Autoencoder MAE 1.82e+06 in Table 4, 1.8e+06 in Table 6, and 2.65e+04 in Table 8), and the values are not monotone in data fraction. The text acknowledges in §Results that some pretrained networks 'seem to have difficulty converging,' but the tables and Figure 4 include these runs in the reported metrics. This makes the quantitative transfer-learning claims hard to interpret. Please either exclude non-converged runs and state the exclusion criterion, or report them separately and revisit the claim that 'most pretraining tasks are helpful.'
minor comments (5)
  1. [Abstract] The phrase 'labeled datavia transfer learning' is missing a space; it should read 'labeled data via transfer learning.'
  2. [Figure 1(d)] The caption says 'three possible permutations of the 10 nearest neighbors,' while the text and architecture use the 20 nearest neighbors; align the number or clarify the relationship.
  3. [Methods, Eq. (1)] The geometric product ⃗r_i ⃗r_j is not defined for readers who are not already familiar with geometric algebra; a one-sentence definition or a reference to [21] would improve accessibility.
  4. [Abstract and Methods] The abstract says the tasks require no human intervention in labeling, but frame classification uses structure type and temperature as labels; clarify that these labels are automatically available from the data source rather than human-annotated.
  5. [Appendix F, Table 9] The autoencoder embedding has dimensionality 8 while all other learned embeddings have 32; a brief explanation of this choice would help the reader interpret Table 1.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the headline LA/LB discrimination is a zero-shot transfer evaluation, and no target quantity is fitted into the pretraining objectives.

full rationale

The central empirical claim is that a frame-classification network pretrained on Gaussian-perturbed ideal AFLOW unit cells produces embeddings that separate two pre-crystallization liquids (LA/LB, ROC AUC 0.962 in Table 1), beating Steinhardt Q (0.506), local spherical harmonics (0.505), and radial-distance baselines. Nothing in the training pipeline uses the LA/LB comparison or the liquid snapshots: the frame-classification labels are crystal structure types from AFLOW (Results, "We first train a network on the frame classification task with crystal structure prototype data"), and the other self-supervised tasks use denoising, bond masking, bond regression, shift regression, and autoencoding losses whose labels are defined independently of the downstream pair. The evaluation is a zero-shot transfer of the pretrained embeddings to MD snapshots, scored by PCA projection and ROC AUC; the same projection procedure is applied uniformly to all baseline featurizations, so any optimism from the PCA step is not a trained parameter being renamed as a prediction. The self-citations are not load-bearing in the circularity sense: [21] supplies geometric algebra attention layers whose equations are reproduced in Equation 1 and whose code is provided; [28] is only a baseline method; [36] supplies the isotropic pair potentials for the MD trajectories. None is invoked as a uniqueness theorem or as external proof that the LA/LB result holds. Appendix F's remark that the endpoint solids cF4-Cu and hP2-Mg "have identical representations when using nearest-neighbor graphs" is an empirical limitation of the 20-neighbor graph descriptor and motivates concern about single-snapshot robustness, but it does not show that the liquid separation is constructed from the target labels. Concerns about missing error bars, single cooling trajectories, and possible confounds (effective temperature, density, finite-size fluctuations) are correctness/robustness risks, not circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No physical constants are fitted to the target result. The central claim depends on modeling choices (20 neighbors, noise levels, architecture size), on the assumption that AFLOW-with-noise transfers to MD liquids, and on the representativeness of the chosen MD trajectories. No invented physical entities are introduced. The absence of error bars in Table 1 means the empirical claim is a point estimate over these choices.

free parameters (4)
  • Neighbor count K = 20
    Input point clouds are fixed to the 20 nearest neighbors per particle; the representational capacity and the FCC/HCP second-shell distinction depend on this hand-chosen cutoff.
  • Noise scale for SSL perturbations = 0.5 sigma0
    Gaussian noise added in denoising, shift, and noisy bond tasks is fixed at standard deviation 0.5 sigma0; task difficulty, and therefore what representations learn, depends on this choice.
  • Ideal-structure noise levels = 0.01, 0.03, 0.05 sigma0
    Gaussian noise applied to AFLOW unit cell coordinates to emulate thermal disorder and sample neighbor permutations; chosen without a systematic study.
  • Architecture dimensions = N=3 layers, hidden 64, invariant embedding 32, autoencoder bottleneck 8
    Core depth, MLP widths, and bottleneck size are hand-selected; they affect transfer learning and downstream AUC.
assumptions (5)
  • domain assumption A fixed 20-nearest-neighbor local environment is sufficient to characterize relevant order, including differences between FCC and HCP stacking.
    All models and baselines use 20 nearest neighbors; if the LA/LB difference lives mostly outside this shell, the comparison would miss it. Appendix F's claim that cF4-Cu and hP2-Mg have identical nearest-neighbor graph representations makes this assumption especially load-bearing.
  • domain assumption Gaussian-perturbed AFLOW unit cells are a sufficient proxy for the local environments encountered in MD self-assembly liquids and quasicrystals.
    Models are trained only on idealized prototypes with coordinate noise, then evaluated zero-shot on MD liquids; generalization is assumed rather than derived.
  • domain assumption The MD trajectories used for evaluation are representative of the target physical systems.
    Single-component systems of 4096 particles with OPP/LJG potentials cooled over 1e8 timesteps; no replicate trajectories are used for the Table 1 snapshot evaluations.
  • domain assumption Frame classification labels (AFLOW structure prototypes) are a meaningful ground truth for structural order.
    The self-supervised frame classification task trains on structure-type labels from AFLOW; its embedding quality is then judged on simulated liquids. If prototype labels do not reflect structural similarity, the transfer argument weakens.
  • standard math Standard linear-algebra and geometric-algebra identities used by the architecture are correct as applied.
    Equation 1's equivariance claims rely on standard geometric algebra and attention constructions; no machine-checked proof is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Learning for Ordered Three-Dimensional Structures." pith.science (2026). https://pith.science/paper/5JFDCMYV

@misc{pith2026241114680,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Learning for Ordered Three-Dimensional Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5JFDCMYV}},
  note         = {Machine review of arXiv:2411.14680}
}
read the original abstract

Recent work has proven that training large language models with self-supervised tasks and fine-tuning these models to complete new tasks in a transfer learning setting is a powerful idea, enabling the creation of models with many parameters, even with little labeled data; however, the number of domains that have harnessed these advancements has been limited. In this work, we formulate a set of geometric tasks suitable for the large-scale study of ordered three-dimensional structures, without requiring any human intervention in data labeling. We build deep rotation- and permutation-equivariant neural networks based on geometric algebra and use them to solve these tasks on both idealized and simulated three-dimensional structures. Quantifying order in complex-structured assemblies remains a long-standing challenge in materials physics; these models can elucidate the behavior of real self-assembling systems in a variety of ways, from distilling insights from learned tasks without further modification to solving new tasks with smaller amounts of labeled data via transfer learning.

Figures

Figures reproduced from arXiv: 2411.14680 by the authors.

Figure 1
Figure 1. Two types of ordered structural data used in this work. (a) Periodic three-dimensional systems can [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architectures for equivariant, self-supervised transfer learning tasks. (a) The core shared by [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. We use a typical classification MLP on top of a permutation-invariant attention reduction over all [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 3
Figure 3. Figure 3: (a) Two-dimensional projection of embeddings for a set of self-assembled structures using a [PITH_FULL_IMAGE:figures/full_fig_p008_3.png]
Figure 4
Figure 4. Figure 4: Transfer learning performance of rotation-equivariant deep neural networks. Models are first [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: (a) Histogram of predicted frames over a simulated cooling trajectory of particles forming the [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

    arXiv 2005

  2. [2]

    BERT: Pre-training of deep bidirectional transformers for language understanding.arXiv:1810.04805 [cs], October 2018

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding.arXiv:1810.04805 [cs], October 2018

    arXiv 2018

  3. [3]

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri, Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan, Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian...

    work page 2021

  4. [4]

    Jukebox: A generative model for music.CoRR, abs/2005.00341, 2020

    Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music.CoRR, abs/2005.00341, 2020

    arXiv 2005

  5. [5]

    Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, July 2021

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing, July 2021

    work page 2021

  6. [6]

    Multitask prompted training enables zero-shot task generalization

    Victor Sanh, Albert Webson, Colin Raffel, Stephen Bach, Lintang Sutawika, Zaid Alyafeai, Antoine Chaffin, Arnaud Stiegler, Arun Raja, Manan Dey, M Saiful Bari, Canwen Xu, Urmish Thakker, Shanya Sharma Sharma, Eliza Szczechla, Taewoon Kim, Gunjan Chhablani, Nihal Nayak, Debajyoti Datta, Jonathan Chang, Mike Tian-Jian Jiang, Han Wang, Matteo Manica, Sheng S...

    2022

  7. [7]

    Self-supervised learning of local features in 3D point clouds

    Ali Thabet, Humam Alwassel, and Bernard Ghanem. Self-supervised learning of local features in 3D point clouds. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPR W), pages 4048–4052, Seattle, WA, USA, June 2020. IEEE. ISBN 978-1-72819-360-1

    work page 2020

  8. [8]

    Self-supervised few-shot learning on point clouds

    Charu Sharma and Manohar Kaul. Self-supervised few-shot learning on point clouds. InAdvances in Neural Information Processing Systems , volume 33, pages 7212–7221. Curran Associates, Inc., 2020

    work page 2020

Show all 46 references
  1. [9]

    Self-supervised learning on 3D point clouds by learning discrete generative models

    Benjamin Eckart, Wentao Yuan, Chao Liu, and Jan Kautz. Self-supervised learning on 3D point clouds by learning discrete generative models. In2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8244–8253, Nashville, TN, USA, June 2021. IEEE. ISBN 9...

    2021

  2. [10]

    Yatian Pang, Wenxiao Wang, Francis E. H. Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. March 2022

    2022

  3. [11]

    Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence , 4(3):279–287, March 2022

    Yuyang Wang, Jianren Wang, Zhonglin Cao, and Amir Barati Farimani. Molecular contrastive learning of representations via graph neural networks.Nature Machine Intelligence , 4(3):279–287, March 2022

    2022

  4. [12]

    3D infomax improves GNNs for molecular property prediction

    Hannes Stärk, Dominique Beaini, Gabriele Corso, Prudencio Tossou, Christian Dallago, Stephan Günnemann, and Pietro Lió. 3D infomax improves GNNs for molecular property prediction. In Proceedings of the 39th International Conference on Machine Learning , pages 20479–20502. PMLR...

    2022

  5. [13]

    Geometry-enhanced molecular representation learning for property prediction

    Xiaomin Fang, Lihang Liu, Jieqiong Lei, Donglong He, Shanzhuo Zhang, Jingbo Zhou, Fan Wang, Hua Wu, and Haifeng Wang. Geometry-enhanced molecular representation learning for property prediction. Nature Machine Intelligence , 4(2):127–134, February 2022

    2022

  6. [14]

    Molecular Geometry Pretraining with SE(3)-Invariant Denoising Distance Matching

    Shengchao Liu, Hongyu Guo, and Jian Tang. Molecular Geometry Pretraining with SE(3)-Invariant Denoising Distance Matching. InThe Eleventh International Conference on Learning Representations , February 2023

    2023

  7. [15]

    Crystal diffusion variational autoencoder for periodic material generation

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crystal diffusion variational autoencoder for periodic material generation. In International Conference on Learning Representations, page 20, 2022. 11

    2022

  8. [16]

    Crystal twins: Self-supervised learning for crystalline material property prediction, May 2022

    Rishikesh Magar, Yuyang Wang, and Amir Barati Farimani. Crystal twins: Self-supervised learning for crystalline material property prediction, May 2022

    2022

  9. [17]

    Mehl, David Hicks, Cormac Toher, Ohad Levy, Robert M

    Michael J. Mehl, David Hicks, Cormac Toher, Ohad Levy, Robert M. Hanson, Gus Hart, and Stefano Curtarolo. The AFLOW library of crystallographic prototypes: Part 1.Computational Materials Science, 136:S1–S828, August 2017

    2017

  10. [18]

    Mehl, Eric Gossett, Cormac Toher, Ohad Levy, Robert M

    David Hicks, Michael J. Mehl, Eric Gossett, Cormac Toher, Ohad Levy, Robert M. Hanson, Gus Hart, and Stefano Curtarolo. The AFLOW library of crystallographic prototypes: Part 2.Computational Materials Science, 161:S1–S1011, April 2019

    2019

  11. [19]

    Howard, Joshua A

    Michael P. Howard, Joshua A. Anderson, Arash Nikoubashman, Sharon C. Glotzer, and Athanassios Z. Panagiotopoulos. Efficient neighbor list calculation for molecular simulation of colloidal systems using graphics processing units. Computer Physics Communications , 203:45–52, June 2016

    2016

  12. [20]

    Anderson, Jens Glaser, and Sharon C

    Joshua A. Anderson, Jens Glaser, and Sharon C. Glotzer. HOOMD-blue: A Python package for high- performance molecular dynamics and hard particle Monte Carlo simulations.Computational Materials Science, 173:109363, 2020

    2020

  13. [21]

    Geometric algebra attention networks for small point clouds.arXiv:2110.02393 [cs], October 2021

    Matthew Spellings. Geometric algebra attention networks for small point clouds.arXiv:2110.02393 [cs], October 2021

    2021 arXiv

  14. [22]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In Yoshua Bengio and Yann LeCun, editors, 2nd International Conference on Learning Representations (ICLR) , 2014

    2014

  15. [23]

    Why does unsupervised pre-training help deep learning? page 36, 2010

    Dumitru Erhan, Yoshua Bengio, Aaron Courville, Pierre-Antoine Manzagol, Pascal Vincent, and Samy Bengio. Why does unsupervised pre-training help deep learning? page 36, 2010

    2010

  16. [24]

    Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion

    Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, and Pierre-Antoine Manzagol. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. page 38, 2010

    2010

  17. [25]

    On lines and planes of closest fit to systems of points in space.The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science , 2(11):559–572, 1901

    Karl Pearson. On lines and planes of closest fit to systems of points in space.The London, Edinburgh, and Dublin Philosophical Magazine and Journal of Science , 2(11):559–572, 1901

    1901

  18. [26]

    J. A. Hanley and B. J. McNeil. A method of comparing the areas under receiver operating characteristic curves derived from the same cases.Radiology, 148(3):839–843, September 1983

    1983

  19. [27]

    Steinhardt, David R

    Paul J. Steinhardt, David R. Nelson, and Marco Ronchetti. Bond-orientational order in liquids and glasses. Physical Review B, 28(2):784–805, July 1983

    1983

  20. [28]

    Machine learning for crystal identification and discovery

    Matthew Spellings and Sharon C Glotzer. Machine learning for crystal identification and discovery. AIChE Journal, 64(6):2198–2206, 2018

    2018

  21. [29]

    Revealing the role of liquid preordering in crystallisation of supercooled liquids

    Yuan-Chao Hu and Hajime Tanaka. Revealing the role of liquid preordering in crystallisation of supercooled liquids. Nature Communications, 13(1):4519, August 2022

    2022

  22. [30]

    Coifman and Stéphane Lafon

    Ronald R. Coifman and Stéphane Lafon. Diffusion maps.Applied and Computational Harmonic Analysis , 21(1):5–30, 2006

    2006

  23. [31]

    Variational inference with normalizing flows

    Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing flows. arXiv:1505.05770 [cs, stat] , May 2015

    2015 arXiv

  24. [32]

    Liquid–liquid phase separation in cellular signaling systems

    P Andrew Chong and Julie D Forman-Kay. Liquid–liquid phase separation in cellular signaling systems. Current Opinion in Structural Biology , 41:180–186, December 2016

    2016

  25. [33]

    Alain A. M. André and Evan Spruijt. Liquid–Liquid phase separation in crowded environments. International Journal of Molecular Sciences , 21(16):5908, January 2020. 12

    2020

  26. [34]

    Nonclassical nucleation pathways in protein crystallization.Journal of Physics

    Fajun Zhang. Nonclassical nucleation pathways in protein crystallization.Journal of Physics. Condensed Matter: An Institute of Physics Journal , 29(44):443002, November 2017

    2017

  27. [35]

    A primer in BERTology: What we know about how BERT works

    Anna Rogers, Olga Kovaleva, and Anna Rumshisky. A primer in BERTology: What we know about how BERT works. Transactions of the Association for Computational Linguistics , 8:842–866, January 2021

    2021

  28. [36]

    Damasceno, Carolyn L

    Julia Dshemuchadse, Pablo F. Damasceno, Carolyn L. Phillips, Michael Engel, and Sharon C. Glotzer. Moving beyond the constraints of chemistry via crystal structure discovery with isotropic multiwell pair potentials. Proceedings of the National Academy of Sciences , 118(21), May 2021

    2021

  29. [37]

    Unsupervised learning for local structure detection in colloidal systems.The Journal of Chemical Physics , 151(15):154901, October 2019

    Emanuele Boattini, Marjolein Dijkstra, and Laura Filion. Unsupervised learning for local structure detection in colloidal systems.The Journal of Chemical Physics , 151(15):154901, October 2019

    2019

  30. [38]

    Kapfer, Gerd E

    Walter Mickel, Sebastian C. Kapfer, Gerd E. Schröder-Turk, and Klaus Mecke. Shortcomings of the bond orientational order parameters for the analysis of disordered particulate matter.The Journal of Chemical Physics, 138(4):044501, January 2013

    2013

  31. [39]

    Dzugutov

    M. Dzugutov. Formation of a dodecagonal quasicrystalline phase in a simple monatomic liquid.Physical Review Letters, 70:2924–2927, 1993

    1993

  32. [40]

    body-centered cubic

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Yoshua Bengio and Yann LeCun, editors,3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015. 13 A Frame classif...

    2015

  33. [41]

    This system is a control experiment: identical liquids should be difficult or impossible to distinguish to a significant degree

    LA/LA: Two liquids at the same temperature that will later formcF 4-Cu. This system is a control experiment: identical liquids should be difficult or impossible to distinguish to a significant degree

  34. [42]

    This models a simple phase transition between low-density and high-density disordered phases

    GA/LA: A gas and liquid that will later formcI2-W. This models a simple phase transition between low-density and high-density disordered phases

  35. [43]

    This models a simple disorder-order transition

    LA/SA: A liquid that formscI2-W and cI2 − W solid. This models a simple disorder-order transition

  36. [44]

    This models a disorder-order transition to a complex ordered structure

    GA/SA: A gas that will form an icosahedral quasicrystal and the quasicrystal itself. This models a disorder-order transition to a complex ordered structure

  37. [45]

    These solid structures are very similar, having identical representations when using nearest-neighbor graphs

    LA/LB: Two liquids that will formcF 4-Cu and hP 2-Mg later. These solid structures are very similar, having identical representations when using nearest-neighbor graphs

  38. [46]

    Most methods developed to distinguish solid structures should be able to handle this baseline

    SA/SB: Two solid structures ofcF 4-Cu and hP 2-Mg. Most methods developed to distinguish solid structures should be able to handle this baseline. 17 Table 6: Transfer learning performance for models trained on a source task (indicated by column) for a target task (indicated by...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.