Pith. sign in

REVIEW 5 major objections 8 minor 64 references

UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation

T0 review · 5 major / 8 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A unified model treats mechanical metamaterial design as one three-way inference problem, where any two modalities—topology, density, properties—determine the third.

desk verdict The model and task framing are solid, but the rotation-augmented split likely leaks base topologies into test, so the headline empirical margins are not interpretable as generalization. read the letter →

arxiv 2506.15722 v1 pith:7RE24HFM submitted 2025-06-05 cs.LG cs.AI

Wangzhi Zhan , Jianpeng Chen , Dongqi Fu , Dawei Zhou This is my paper
classification cs.LGcs.AI
keywords mechanicalmetamaterialsunifiedmodeltopologygenerationpropertypredictionconditionconfirmationtripartiteoptimaltransportpartiallyfrozendiffusionlatticestructures
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

This paper argues that mechanical metamaterial design should be treated as one three-way inference problem rather than as separate tasks. The authors introduce a unified model, UNIMATE, that takes any two of the three defining modalities—3D topology $T$, relative density $\rho$, and mechanical properties $p$—and produces the third, covering topology generation, property prediction, and what they call condition confirmation. To make this work, all three modalities are compressed into a shared discrete latent space whose tokens are aligned by a tripartite optimal transport scheme, then completed by a partially frozen score-based diffusion model that keeps the known inputs fixed. On a new benchmark of 15,000 rotation-augmented truss samples, UNIMATE reports lower error than six task-specific baselines on all three tasks, with the largest improvement (80.2%) in topology generation quality. If the results hold, a single model could replace separate generators, property predictors, and density checkers in metamaterial design pipelines.

What carries the argument

The load-bearing mechanism is the Latent MTR token sequence, a discrete codebook representation that unifies the three modalities. Each modality is passed through its own encoder (a graph network for topology, MLPs for density and properties) and rounded to the nearest prototype token; a tripartite optimal transport alignment, implemented by a generalized Sinkhorn iteration, pulls the three token distributions together. The synergetic generation module then runs a score-based diffusion process over the token sequence while freezing known tokens at their initial values, so unknown tokens are generated conditionally on any supplied subset. This combination is what lets a single pipeline switch between generation, prediction, and confirmation by changing only which tokens are masked.

What would settle it

Re-run the benchmark with a topology-wise split—putting all ten rotations of a selected topology in the same fold—and compare condition-confirmation NRMSE and property-prediction NRMSE; if those errors jump substantially while generation metrics stay similar, the original gains were inflated by rotated copies crossing the train/test boundary. A direct check is to search the training set for any test topology under a different rotation angle.

Watch

Extended reading notes

Core claim

UNIMATE's central claim is that the three modalities of a mechanical metamaterial—topology $T$, density condition $\rho$, and mechanical properties $p$—can be embedded into one token sequence, the Latent MTR, such that completing missing tokens by diffusion solves all three design tasks at once. The authors formalize the Metamaterial Trinity Representation $M=(T,\rho,p)$ and show that knowing any two entries should determine the third. Their experiments report that UNIMATE beats baseline models on every task, with improvements over the best baseline of up to 80.2% in topology generation, 5.1% in property prediction, and 50.2% in condition confirmation. They also present a benchmark covering all three modalities, built by selecting 500 topologies, assigning three densities to each, computing properties by homogenization simulation, and rotating each sample nine times.

Load-bearing premise

The results assume the random 70/15/15 split of the rotation-augmented data keeps all rotated copies of one topology in the same split; if the split is per-sample, the same topology can appear in both training and test, and the reported property and density errors could come partly from memorizing duplicates rather than generalizing to new topologies.

Editorial extensions

If this is right

  • One trained UNIMATE model can act as an inverse design tool: given a target stiffness and a density budget, it generates a 3D topology; given a topology and density, it predicts the full property vector; given a topology and target properties, it confirms the needed density.
  • Because the synergetic generation module accepts an arbitrary subset of known tokens, the same machinery covers any partial specification of the trinity, not just the three canonical tasks.
  • The new 15,000-sample benchmark with homogenization-computed properties gives later work a common yardstick for evaluating all three tasks simultaneously.
  • The ablation study indicates that the latent compression, the alignment operation, and the partially frozen diffusion technique each contribute measurable gains, so each module is load-bearing as claimed.
  • The model's memory efficiency—no out-of-memory error at batch size 10,000 where several baselines fail—means it can scale to larger design searches.

Reading between the lines

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

  • Editorial inference: because density is invariant under the rotation augmentation while topology and property rotate together, the evaluation should be re-run with a topology-wise split; if the original split was sample-wise, the condition-confirmation and property errors could be optimistically biased.
  • Editorial inference: the same token-completion scheme may transfer to other 'trinity' design problems, such as molecular structure, solvent, and target property, since the alignment and frozen-diffusion modules are not metamaterial-specific.
  • Editorial inference: the case study's intermediate structures, which the authors say are absent from the training set, suggest the model can interpolate within the design distribution; one can test this directly by asking whether generated intermediates satisfy finite-element-computed properties, not just the learned surrogate.
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

5 major / 8 minor

Summary. The paper formalizes mechanical metamaterial design as three coupled tasks—topology generation from density/property, property prediction from topology/density, and condition (density) confirmation from topology/property—and proposes UniMate, a unified model combining a VQ-VAE-style discrete latent space, tripartite optimal transport alignment, and a partially frozen score-based diffusion module. A new benchmark is constructed from 500 topologies, each with 3 densities and 9 rotation augmentations, yielding 15,000 samples. Experiments compare UniMate against six baselines and report large improvements across all three tasks, with additional ablations, efficiency analysis, and parameter sensitivity studies.

Significance. If the reported results generalize, the paper would make a useful contribution: a single model handling three interlinked design modalities, a new benchmark for the community, and a principled alignment mechanism. The tripartite OT transport plan is used only as a learned prior for diffusion initialization, so I do not see a circularity problem in the method itself. The open-sourced repository and dataset are also positive elements. However, the central empirical claims currently rest on an ambiguous data-split protocol and single-run numbers; until the split granularity and baseline input protocols are clarified, the headline margins (80.2%, 5.1%, 50.2%) cannot be taken as evidence of generalization to new topologies.

major comments (5)
  1. [Appendix A.1 / Section B.1] The paper never states whether the 70/15/15 split is performed by base topology or by individual augmented sample. Because each of the 500 base topologies appears nine times under rotation with the same density and co-rotated properties, a sample-wise split places rotated copies of training topologies in the test set. Relative density is rotation-invariant, so the condition-confirmation task (NRMSEcc) can be solved by retrieving a memorized scalar for a seen topology, and property prediction can be solved by retrieving a memorized tensor; topology generation can likewise reproduce near-copies of training topologies. The reported margins in Table 2 would then reflect memorization rather than generalization to new topologies. Please state the split granularity, and if it was sample-wise, re-run all experiments with a topology-disjoint split and report both sets of numbers.
  2. [Table 2 / Section 4.2] No error bars, confidence intervals, or significance tests are reported for any comparison. All entries are single runs. The property-prediction margin (NRMSEpp 2.44 vs 2.57, about 5%) is small and could easily arise from random initialization; the generation and condition-confirmation margins are large but still need variance estimates. Report mean plus or minus standard deviation over multiple seeds for every task and perform a paired significance test (or equivalent) before claiming superiority.
  3. [Section 4.2] For the property-prediction task, the paper does not state whether the four prediction baselines (Equiformer, ViSNet, MACE-ve, uniTruss) were given the density condition as input, even though the task definition in Section 2.2 is 'Given the topology T and density condition ρ, predict property p.' If the baselines received only topology, they are handicapped by construction, making the 5.1% comparison uninterpretable. Please describe the exact inputs and any input-conditioning modifications for each baseline, or retrain them with density as an additional input feature.
  4. [Section 4.2 / Appendix A.2.1] Fcond (Eq. 26) is a nearest-neighbor distance from the generated vertices to a single ground-truth topology. For conditional topology generation, many topologies may satisfy the same density and property constraints, so this metric penalizes valid alternative designs and rewards reproducing the training sample; combined with rotation augmentation, a generated rotated copy of a training topology would score artificially well. Please complement the metric with a validity check on the generated structure's simulated properties, and report diversity or coverage metrics.
  5. [Section 4.2 / Table 2] The baseline adaptation for condition confirmation is underspecified. The sentence 'we revise the three prediction-oriented models by forcing them to predict density' does not explain the output head, loss function, or training protocol, so the 50.2% improvement may reflect an unfavorable adaptation rather than a genuine advantage. Please provide the exact architecture changes and training details for the adapted baselines, or replace them with task-specific condition-confirmation models.
minor comments (8)
  1. [Section 3.2] The phrase 'inspired by inspired by VQ-VAE' contains a duplicated word and should be corrected.
  2. [Equation (1)] The GCN update \tilde{X}^l = A\sigma(\tilde{X}^{l-1}, A)W^l appears to apply \sigma to two arguments; this is likely a typo and should read \sigma(\tilde{X}^{l-1}) or similar.
  3. [Table 3 / Section 4.5] The abbreviation 'NRMSE dp' is used in Table 3 and the ablation discussion, while Table 2 uses 'NRMSEcc' for the same task; please define the abbreviation and use consistent notation.
  4. [Figure 4 / Section 4.3] The legend label 'UniMetaMate' should be 'UniMate', and the figure omits baselines that appear in Table C.1; please make the figure consistent with the table.
  5. [Algorithm 1] Line 6 writes 'FT ,k ← −FT ,k + 1'; this should be an increment (FT,k ← FT,k + 1), and the symbols n and lT are not defined in the algorithm block.
  6. [Section 4.4] The codebook size is denoted κ in Section 3.2 but called n in Section 4.4; please align the notation.
  7. [Appendix B.1] Please report the number of diffusion steps, transformer depth, learning rate, batch size, and number of seeds to make the experiments reproducible beyond the released code.
  8. [Section 4.2] The claim of 'up to 80.2%' improvement in topology generation should specify the aggregation; it appears to be the average of the relative improvements in Fqua and Fcond rather than a single-metric improvement, and this should be stated explicitly.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity in the model derivation; the flagged item is an evaluation-support gap: the paper never states whether the 70/15/15 split of the 9x-rotated dataset is topology-disjoint, so the reported NRMSE and Fcond gains could partly reflect train/test overlap rather than generalization to unseen topologies.

  1. fitted input called prediction [Appendix A.1 (data augmentation) and Appendix B.1 (dataset division); governs the Table 2 effectiveness claims.]
    "Based on these 1500 data points, we apply data augmentation by rotating the topology and the properties with the same random rotating angle. Each data point is rotated 9 times, so in total, our dataset contains 15000 samples. ... We use the same dataset division ratio (i.e.,70% for training, 15% for validation and 15% for testing) for each model."

    The paper never states whether the 70/15/15 division is performed per base topology or per augmented sample. Rotation leaves the scalar density unchanged and co-rotates the property tensor, so under a sample-wise split every test sample has nine rotated siblings with identical density in the training set. Condition confirmation then reduces to retrieving the memorized rotation-invariant density of an already-seen topology, and property prediction reduces to retrieving a co-rotated tensor; the reported 50.2% (NRMSEcc) and 5.1% (NRMSEpp) margins would be explained by train/test overlap rather than by generalization to new topologies.

full rationale

UNIMATE's derivation chain is self-contained against external inputs. The three modalities (T, ρ, p) are encoded to discrete tokens (Eqs. 1-6), aligned via tripartite optimal transport (Eqs. 7-9), and completed by a partially frozen score-based diffusion (Eqs. 10-12); the training losses (Eqs. 13-14) are reconstruction and alignment objectives in the LMTR space. No evaluation metric (Fqua/Fsym/Fper/Fcond, NRMSEpp, NRMSEcc, Eqs. 17-28) appears in any training objective, so the reported numbers are not restatements of fitted losses. The TOT transport plan (Alg. 1) is computed from training-token frequencies and only initializes unknown tokens; it is a learned prior, not a parameter that encodes test answers. The model rests on external building blocks: dataset from Lumpe & Stankovic (2021), homogenization simulation (Silveirinha, 2007), codebook from VQ-VAE (Van Den Oord et al., 2017), diffusion from Song & Ermon (2019), and baselines from other groups. The group's own prior work (Metamat-bench, Chen et al., 2025b; Meta-scientist, Qi et al., 2024) appears in the reference list but is not invoked in the body to justify the model or to rule out alternatives, so there is no load-bearing self-citation and no imported uniqueness theorem. The one flagged step is conditional: Appendix A.1 creates 15,000 samples by nine rotations of 1,500 base samples, and Appendix B.1 specifies only the 70/15/15 ratio, never the granularity of the split. If the division is sample-wise, each test sample's rotated siblings (identical scalar density, co-rotated property tensor) sit in the training set, and condition confirmation and property prediction become memorization retrieval; the 80.2%, 5.1%, and 50.2% margins would then not measure transfer to unseen topologies. This is a missing-statement and support gap rather than an algebraic circularity: no equation of the paper forces the reported outputs from the inputs, and a stated topology-wise split would resolve it. Also flagged as missing support: Section 3.2 refers lattice-vector computation to "the method mentioned in Appendix A.2," but Appendix A.2 defines only the evaluation metrics and contains no lattice-vector method. These items belong to correctness and verification risk and keep the circularity score at 2.

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

The model relies on standard ML training plus several domain assumptions about simulation fidelity, metric validity, augmentation correctness, and the absence of data leakage. No new physical entities are introduced. The free parameters are mostly hyperparameters that are not fully reported, which weakens reproducibility and makes the exact experimental configuration unclear.

free parameters (4)
  • Alignment loss weights α_T, α_ρ, α_p, α_w = not reported
    Equation 9 combines reconstruction, rounding, and transport losses with four hyperparameters whose values are not specified in the paper.
  • Total loss weights λ_align, λ_gen = not reported
    Equation 14 balances alignment and generation losses; the chosen values are not given.
  • Codebook size κ and latent token dimension d = not reported
    Sensitivity analysis in Figure 5 and Table C.2 varies these, but the final configuration used for the main results is not stated.
  • Number of densities per topology (3) and rotation augmentations (9) = 3 and 9
    Dataset construction choices in Appendix A.1; the narrow density coverage makes condition confirmation substantially easier and limits the generality of that result.
assumptions (4)
  • domain assumption Homogenization simulation (Silveirinha 2007) yields accurate mechanical properties for the voxel-meshed unit cells.
    Properties used as ground truth in training and evaluation; any simulation error propagates directly into all three task metrics (Appendix A.1).
  • domain assumption The proposed Fqua and Fcond metrics reliably measure topology quality and condition matching.
    These are new metrics introduced in Appendix A.2 with no validation against physical simulation or human judgment; the headline 80.2% improvement is computed on them.
  • domain assumption Rotating topology and properties jointly is a valid augmentation that preserves the topology-property mapping.
    Appendix A.1 states the augmentation but does not specify how the property tensor is transformed; if rotated incorrectly, the augmented dataset is internally inconsistent.
  • domain assumption Random sample-wise split of the augmented data does not create train/test leakage.
    Section B.1 states only the split ratio; if the same topology appears in both train and test under different rotations, condition confirmation and property prediction results are inflated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation." pith.science (2026). https://pith.science/paper/7RE24HFM

@misc{pith2026250615722,
  author       = {Pith},
  title        = {Pith review of: UniMate: A Unified Model for Mechanical Metamaterial Generation, Property Prediction, and Condition Confirmation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7RE24HFM}},
  note         = {Machine review of arXiv:2506.15722}
}
read the original abstract

Metamaterials are artificial materials that are designed to meet unseen properties in nature, such as ultra-stiffness and negative materials indices. In mechanical metamaterial design, three key modalities are typically involved, i.e., 3D topology, density condition, and mechanical property. Real-world complex application scenarios place the demanding requirements on machine learning models to consider all three modalities together. However, a comprehensive literature review indicates that most existing works only consider two modalities, e.g., predicting mechanical properties given the 3D topology or generating 3D topology given the required properties. Therefore, there is still a significant gap for the state-of-the-art machine learning models capturing the whole. Hence, we propose a unified model named UNIMATE, which consists of a modality alignment module and a synergetic diffusion generation module. Experiments indicate that UNIMATE outperforms the other baseline models in topology generation task, property prediction task, and condition confirmation task by up to 80.2%, 5.1%, and 50.2%, respectively. We opensource our proposed UNIMATE model and corresponding results at https://github.com/wzhan24/UniMate.

Figures

Figures reproduced from arXiv: 2506.15722 by the authors.

Figure 1
Figure 1. Trinity Representation of Metamaterials. offering novel solutions for challenges in engineering, man￾ufacturing, and materials science (Barchiesi et al., 2019). Metamaterials can have many possible important merits that classic materials do not have, like negative Poisson’s ratio (Wang et al., 2020a) (favored for soft device), energy absorption behavior (Yuan et al., 2019) (favored for cush￾ioning device), tunabilit… view at source ↗
Figure 2
Figure 2. Hierarchical Composition of Mechanical Metamaterial. 3D Topology is represented by T = (L, X, A), where L = [l1,l2,l3] contains the three lattice vectors indicating the axes along which the unit cell is stacked; X ∈ R n×3 is the 3D coordinates matrix of nodes, and A ∈ R n×n is the adjacency matrix without self-loop, where n denotes the node numbers, with Ai,j = 1 if ith node connects jth node. Relative Density Condi… view at source ↗
Figure 3
Figure 3. Pipeline Illustration of UNIMATE. Three modalities are fed into the model, crossing two modules, and generating reconstructed data for training. In the inference phase, the targeted modality data is generated from noise. i.e., whether and how well a specific topology can perform under a particular density. Condition Confirmation. Given topology T and property p, we aim to confirm the corresponding density condition … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of Time Efficiency. considered. 4.4. Parameter Sensitivity We also study the parameter sensitivity of our model by tuning the latent token dimension d and number of to￾kens within the codebook n. The resulting generation met￾ric Fqua is shown in [PITH_FULL_…
Figure 5
Figure 5. Figure 5: Fqua under different parameters. 4.5. Ablation Study [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: shows that our model suggests octet truss topology in the HSLD-targeted task, while octet truss is known to be a promising candidate for high-stiffness topology (Song et al., 2019). Our model can also generate some novel intermediate topologies that are not in the trai…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 53 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    Mechanical metamaterials: a state of the art

    Barchiesi, E., Spagnuolo, M., and Placidi, L. Mechanical metamaterials: a state of the art. Mathematics and Mechanics of Solids, 24 0 (1): 0 212--234, 2019

    work page 2019

  3. [3]

    M., Dzamba, M., Gao, M., Rizvi, A., Zitnick, C

    Barroso-Luque, L., Shuaibi, M., Fu, X., Wood, B. M., Dzamba, M., Gao, M., Rizvi, A., Zitnick, C. L., and Ulissi, Z. W. Open materials 2024 (omat24) inorganic materials dataset and models, 2024. URL https://arxiv.org/abs/2410.12771

    arXiv 2024

  4. [4]

    and Kochmann, D

    Bastek, J.-H. and Kochmann, D. M. Inverse design of nonlinear mechanical metamaterials via video denoising diffusion models. Nature Machine Intelligence, 5 0 (12): 0 1466--1475, 2023

    work page 2023

  5. [5]

    N., and Kochmann, D

    Bastek, J.-H., Kumar, S., Telgen, B., Glaesener, R. N., and Kochmann, D. M. Inverting the structure--property map of truss metamaterials by deep learning. Proceedings of the National Academy of Sciences, 119 0 (1): 0 e2111505119, 2022 a

    work page 2022

  6. [6]

    N., and Kochmann, D

    Bastek, J.-H., Kumar, S., Telgen, B., Glaesener, R. N., and Kochmann, D. M. Inverting the structure--property map of truss metamaterials by deep learning. Proceedings of the National Academy of Sciences, 119 0 (1): 0 e2111505119, 2022 b

    work page 2022

  7. [7]

    P., Simm, G

    Batatia, I., Kovacs, D. P., Simm, G. N. C., Ortner, C., and Csanyi, G. MACE : Higher order equivariant message passing neural networks for fast and accurate force fields. In Oh, A. H., Agarwal, A., Belgrave, D., and Cho, K. (eds.), Advances in Neural Information Processing Systems, 2022. URL https://openreview.net/forum?id=YPpSngE-ZU

    work page 2022

  8. [8]

    S., and He, L

    Chen, J., Ling, Y., Xu, J., Ren, Y., Huang, S., Pu, X., Hao, Z., Yu, P. S., and He, L. Variational graph generator for multiview graph clustering. IEEE Transactions on Neural Networks and Learning Systems, pp.\ 1--14, 2025 a

    work page 2025

Show all 64 references
  1. [9]

    Metamatbench: Integrating heterogeneous data, computational tools, and visual interface for metamaterial discovery

    Chen, J., Zhan, W., Wang, H., Jia, Z., Gan, J., Zhang, J., Qi, J., Chen, T., Huang, L., Chen, M., et al. Metamatbench: Integrating heterogeneous data, computational tools, and visual interface for metamaterial discovery. arXiv preprint arXiv:2505.20299, 2025 b

    2025 arXiv

  2. [10]

    J., Yildirim, B., Jain, A., and Cole, J

    Court, C. J., Yildirim, B., Jain, A., and Cole, J. M. 3-d inorganic crystal structure generation and property prediction via representation learning. Journal of Chemical Information and Modeling, 60 0 (10): 0 4518--4535, 2020

    2020

  3. [11]

    C., Zhu, Y., Xie, T., Duan, C., Gregoire, J., and Gomes, C

    Du, Y., Wang, Y., Huang, Y., Li, J. C., Zhu, Y., Xie, T., Duan, C., Gregoire, J., and Gomes, C. P. M ^ 2 hub: Unlocking the potential of machine learning for materials discovery. Advances in Neural Information Processing Systems, 36: 0 77359--77378, 2023

    2023

  4. [12]

    Benchmarking materials property prediction methods: the matbench test set and automatminer reference algorithm

    Dunn, A., Wang, Q., Ganose, A., Dopp, D., and Jain, A. Benchmarking materials property prediction methods: the matbench test set and automatminer reference algorithm. npj Computational Materials, 6 0 (1): 0 138, 2020

    2020

  5. [13]

    and Ziolkowski, R

    Engheta, N. and Ziolkowski, R. W. Metamaterials: physics and engineering explorations. John Wiley & Sons, 2006

    2006

  6. [14]

    Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules

    Gebauer, N., Gastegger, M., and Sch \"u tt, K. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. Advances in neural information processing systems, 32, 2019

    2019

  7. [15]

    u ller, K.-R., and Sch \

    Gebauer, N. W., Gastegger, M., Hessmann, S. S., M \"u ller, K.-R., and Sch \"u tt, K. T. Inverse design of 3d molecular structures with conditional generative neural networks. Nature communications, 13 0 (1): 0 973, 2022

    2022

  8. [16]

    Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials

    Grega, I., Batatia, I., Csanyi, G., Karlapati, S., and Deshpande, V. Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials. In ICLR, 2024. URL https://openreview.net/forum?id=smy4DsUbBo

    2024

  9. [17]

    G., Zitnick, C

    Gruver, N., Sriram, A., Madotto, A., Wilson, A. G., Zitnick, C. L., and Ulissi, Z. W. Fine-tuned language models generate stable inorganic materials as text. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=vN9fpfqoP1

    2024

  10. [18]

    G., Vignac, C., and Welling, M

    Hoogeboom, E., Satorras, V. G., Vignac, C., and Welling, M. Equivariant diffusion for molecule generation in 3d. In ICML, pp.\ 8867--8887. PMLR, 2022

    2022

  11. [19]

    P., Karlapati, S., Shaikeea, A

    Indurkar, P. P., Karlapati, S., Shaikeea, A. J. D., and Deshpande, V. S. Predicting deformation mechanisms in architected metamaterials using gnn. arXiv preprint arXiv:2202.09427, 2022

    2022 arXiv

  12. [20]

    Designing three-dimensional lattice structures with anticipated properties through a deep learning method

    Jia, Z., Gong, H., Liu, S., Zhang, J., and Zhang, Q. Designing three-dimensional lattice structures with anticipated properties through a deep learning method. Materials & Design, 244: 0 113139, 2024. ISSN 0264-1275. doi:https://doi.org/10.1016/j.matdes.2024.113139. URL https:...

    2024

  13. [21]

    Crystal structure prediction by joint equivariant diffusion on lattices and fractional coordinates

    Jiao, R., Huang, W., Lin, P., Han, J., Chen, P., Lu, Y., and Liu, Y. Crystal structure prediction by joint equivariant diffusion on lattices and fractional coordinates. In Workshop on ''Machine Learning for Materials'' ICLR 2023, 2023. URL https://openreview.net/forum?id=VPByphdu24j

    2023

  14. [22]

    Kingma, D. P. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  15. [23]

    Kipf, T. N. and Welling, M. Variational graph auto-encoders. arXiv preprint arXiv:1611.07308, 2016

    2016 arXiv

  16. [24]

    Lee, S., Zhang, Z., and Gu, G. X. Generative machine learning algorithm for lattice structures with superior mechanical properties. Materials Horizons, 9 0 (3): 0 952--960, 2022

    2022

  17. [25]

    and Smidt, T

    Liao, Y.-L. and Smidt, T. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. arXiv preprint arXiv:2206.11990, 2022

    2022 arXiv

  18. [26]

    Equivariant diffusion for crystal structure prediction

    Lin, P., Chen, P., Jiao, R., Mo, Q., Jianhuan, C., Huang, W., Liu, Y., Huang, D., and Lu, Y. Equivariant diffusion for crystal structure prediction. In Forty-first International Conference on Machine Learning, 2024

    2024

  19. [27]

    Symmetry-informed geometric representation for molecules, proteins, and crystalline materials

    Liu, S., Li, Y., Li, Z., Zheng, Z., Duan, C., Ma, Z.-M., Yaghi, O., Anandkumar, A., Borgs, C., Chayes, J., et al. Symmetry-informed geometric representation for molecules, proteins, and crystalline materials. Advances in neural information processing systems, 36, 2024

    2024

  20. [28]

    Spherical message passing for 3d molecular graphs

    Liu, Y., Wang, L., Liu, M., Lin, Y., Zhang, X., Oztekin, B., and Ji, S. Spherical message passing for 3d molecular graphs. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=givsRXsOt9r

    2022

  21. [29]

    Lumpe, T. S. and Stankovic, T. Exploring the property space of periodic cellular structures based on crystal networks. Proceedings of the National Academy of Sciences, 118 0 (7): 0 e2003504118, 2021

    2021

  22. [30]

    Deep learning generative model for crystal structure prediction

    Luo, X., Wang, Z., Gao, P., Lv, J., Wang, Y., Chen, C., and Ma, Y. Deep learning generative model for crystal structure prediction. npj Computational Materials, 10 0 (1): 0 254, 2024 a

    2024

  23. [31]

    Towards symmetry-aware generation of periodic materials

    Luo, Y., Liu, C., and Ji, S. Towards symmetry-aware generation of periodic materials. Advances in Neural Information Processing Systems, 36, 2024 b

    2024

  24. [32]

    Predicting stress, strain and deformation fields in materials and structures with graph neural networks

    Maurizi, M., Gao, C., and Berto, F. Predicting stress, strain and deformation fields in materials and structures with graph neural networks. Scientific reports, 12 0 (1): 0 21834, 2022

    2022

  25. [33]

    P., Bonatti, C., Tancogne-Dejean, T., and Mohr, D

    Meyer, P. P., Bonatti, C., Tancogne-Dejean, T., and Mohr, D. Graph-based metamaterials: Deep learning of structure-property relations. Materials & Design, 223: 0 111175, 2022

    2022

  26. [34]

    A., Bengio, Y., Sharma, D., Carrier, P

    Mistal, Hern \'a ndez-Garc \' a, A., Volokhova, A., Duval, A. A., Bengio, Y., Sharma, D., Carrier, P. L., Koziarski, M., and Schmidt, V. Crystal- GFN : sampling materials with desirable properties and constraints. In AI for Accelerated Materials Design - NeurIPS 2023 Workshop,...

    2023

  27. [35]

    A., Ramsden, S

    O’keeffe, M., Peskov, M. A., Ramsden, S. J., and Yaghi, O. M. The reticular chemistry structure resource (rcsr) database of, and symbols for, crystal nets. Accounts of chemical research, 41 0 (12): 0 1782--1789, 2008

    2008

  28. [36]

    D., et al

    Qi, J., Jia, Z., Liu, M., Zhan, W., Zhang, J., Wen, X., Gan, J., Chen, J., Liu, Q., Ma, M. D., et al. Metascientist: A human-ai synergistic framework for automated mechanical metamaterial design. arXiv preprint arXiv:2412.16270, 2024

    2024 arXiv

  29. [37]

    Three-dimensional euclidean nets from two-dimensional hyperbolic tilings: kaleidoscopic examples

    Ramsden, S., Robins, V., and Hyde, S. Three-dimensional euclidean nets from two-dimensional hyperbolic tilings: kaleidoscopic examples. Acta Crystallographica Section A: Foundations of Crystallography, 65 0 (2): 0 81--108, 2009

    2009

  30. [38]

    Inverse design of crystals using generalized invertible crystallographic representation

    Ren, Z., Noh, J., Tian, S., Oviedo, F., Xing, G., Liang, Q., Aberle, A., Liu, Y., Li, Q., Jayavelu, S., et al. Inverse design of crystals using generalized invertible crystallographic representation. arXiv preprint arXiv:2005.07609, 3 0 (6): 0 7, 2020

    2005 arXiv

  31. [39]

    u tt, K., Kindermans, P.-J., Sauceda Felix, H. E., Chmiela, S., Tkatchenko, A., and M \

    Sch \"u tt, K., Kindermans, P.-J., Sauceda Felix, H. E., Chmiela, S., Tkatchenko, A., and M \"u ller, K.-R. Schnet: A continuous-filter convolutional neural network for modeling quantum interactions. Advances in neural information processing systems, 30, 2017

    2017

  32. [40]

    Equivariant message passing for the prediction of tensorial properties and molecular spectra

    Sch \"u tt, K., Unke, O., and Gastegger, M. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning, pp.\ 9377--9388. PMLR, 2021

    2021

  33. [41]

    and Varadan, V

    Sheng, Z. and Varadan, V. V. Tuning the effective properties of metamaterials by changing the substrate properties. Journal of applied physics, 101 0 (1), 2007

    2007

  34. [42]

    Silveirinha, M. G. Metamaterial homogenization approach with application to the characterization of microstructured composites with negative parameters. Physical Review B—Condensed Matter and Materials Physics, 75 0 (11): 0 115104, 2007

    2007

  35. [43]

    and Mukhopadhyay, T

    Sinha, P. and Mukhopadhyay, T. Programmable multi-physical mechanics of mechanical metamaterials. Materials Science and Engineering: R: Reports, 155: 0 100745, 2023

    2023

  36. [44]

    Diagonal equivalence to matrices with prescribed row and column sums

    Sinkhorn, R. Diagonal equivalence to matrices with prescribed row and column sums. The American Mathematical Monthly, 74 0 (4): 0 402--405, 1967

    1967

  37. [45]

    Octet-truss cellular materials for improved mechanical properties and specific energy absorption

    Song, J., Zhou, W., Wang, Y., Fan, R., Wang, Y., Chen, J., Lu, Y., and Li, L. Octet-truss cellular materials for improved mechanical properties and specific energy absorption. Materials & Design, 173: 0 107773, 2019

    2019

  38. [46]

    and Ermon, S

    Song, Y. and Ermon, S. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019

    2019

  39. [47]

    U., Gao, L., Du, H., Li, X., Xiong, X., Fang, N

    Surjadi, J. U., Gao, L., Du, H., Li, X., Xiong, X., Fang, N. X., and Lu, Y. Mechanical metamaterials and their engineering applications. Advanced Engineering Materials, 21 0 (3): 0 1800864, 2019

    2019

  40. [48]

    Neural discrete representation learning

    Van Den Oord, A., Vinyals, O., et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017

    2017

  41. [49]

    Vogel, G., Sortino, P., and Weber, J. M. Graph-to-string variational autoencoder for synthetic polymer design. In AI for Accelerated Materials Design-NeurIPS 2023 Workshop, 2023

    2023

  42. [50]

    A novel two-dimensional mechanical metamaterial with negative poisson’s ratio

    Wang, H., Zhang, Y., Lin, W., and Qin, Q.-H. A novel two-dimensional mechanical metamaterial with negative poisson’s ratio. Computational Materials Science, 171: 0 109232, 2020 a

    2020

  43. [51]

    Ultra-low density architectured metamaterial with superior mechanical properties and energy absorption capability

    Wang, Q., Li, Z., Zhang, Y., Cui, S., Yang, Z., and Lu, Z. Ultra-low density architectured metamaterial with superior mechanical properties and energy absorption capability. Composites Part B: Engineering, 202: 0 108379, 2020 b

    2020

  44. [52]

    Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing

    Wang, Y., Wang, T., Li, S., He, X., Li, M., Wang, Z., Zheng, N., Shao, B., and Liu, T.-Y. Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing. Nature Communications, 15 0 (1): 0 313, 2024

    2024

  45. [53]

    and Evans, R

    Wortman, J. and Evans, R. Young's modulus, shear modulus, and poisson's ratio in silicon and germanium. Journal of applied physics, 36 0 (1): 0 153--156, 1965

    1965

  46. [54]

    and Grossman, J

    Xie, T. and Grossman, J. C. Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties. Physical review letters, 120 0 (14): 0 145301, 2018

    2018

  47. [55]

    Xie, T., Fu, X., Ganea, O.-E., Barzilay, R., and Jaakkola, T. S. Crystal diffusion variational autoencoder for periodic material generation. In International Conference on Learning Representations, 2022. URL https://openreview.net/forum?id=03RLpj-tc_

    2022

  48. [56]

    Complete and efficient graph transformers for crystal material property prediction

    Yan, K., Fu, C., Qian, X., Qian, X., and Ji, S. Complete and efficient graph transformers for crystal material property prediction. arXiv preprint arXiv:2403.11857, 2024

    2024 arXiv

  49. [57]

    Diffusion models: A comprehensive survey of methods and applications

    Yang, L., Zhang, Z., Song, Y., Hong, S., Xu, R., Zhao, Y., Zhang, W., Cui, B., and Yang, M.-H. Diffusion models: A comprehensive survey of methods and applications. ACM Computing Surveys, 56 0 (4): 0 1--39, 2023

    2023

  50. [58]

    Mechanical metamaterials associated with stiffness, rigidity and compressibility: A brief review

    Yu, X., Zhou, J., Liang, H., Jiang, Z., and Wu, L. Mechanical metamaterials associated with stiffness, rigidity and compressibility: A brief review. Progress in Materials Science, 94: 0 114--173, 2018

    2018

  51. [59]

    K., and Zhou, K

    Yuan, S., Chua, C. K., and Zhou, K. 3d-printed mechanical metamaterials with high energy absorption. Advanced Materials Technologies, 4 0 (3): 0 1800419, 2019

    2019

  52. [60]

    Zheng, L., Karapiperis, K., Kumar, S., and Kochmann, D. M. Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling. Nature Communications, 14 0 (1): 0 7563, 2023 a

    2023

  53. [61]

    Zheng, L., Karapiperis, K., Kumar, S., and Kochmann, D. M. Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling. Nature Communications, 14 0 (1): 0 7563, 2023 b

    2023

  54. [62]

    Fairgen: Towards fair graph generation

    Zheng, L., Zhou, D., Tong, H., Xu, J., Zhu, Y., and He, J. Fairgen: Towards fair graph generation. In 2024 IEEE 40th International Conference on Data Engineering (ICDE), pp.\ 2285--2297. IEEE, 2024

    2024

  55. [63]

    Misc-gan: A multi-scale generative model for graphs

    Zhou, D., Zheng, L., Xu, J., and He, J. Misc-gan: A multi-scale generative model for graphs. Frontiers in big Data, 2: 0 3, 2019

    2019

  56. [64]

    A data-driven graph generative model for temporal interaction networks

    Zhou, D., Zheng, L., Han, J., and He, J. A data-driven graph generative model for temporal interaction networks. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, pp.\ 401--411, 2020

    2020

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.