pith. machine review for the scientific record. sign in

arxiv: 2604.27300 · v1 · submitted 2026-04-30 · 💻 cs.AI

Recognition: unknown

METASYMBO: Multi-Agent Language-Guided Metamaterial Discovery via Symbolic Latent Evolution

Dawei Zhou, Dongqi Fu, Jianpeng Chen, Junkai Zhang, Ling Li, Wangzhi Zhan, Wei Wang, Zian Jia

Authors on Pith no claims yet

Pith reviewed 2026-05-07 08:49 UTC · model grok-4.3

classification 💻 cs.AI
keywords metamaterial discoverylanguage-guided designmulti-agent systemssymbolic latent evolutionmicrostructure generationinverse designstructural validitygenerative models
0
0 comments X

The pith

A multi-agent framework with symbolic latent evolution generates valid metamaterial structures directly from natural language intents.

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

Metamaterial discovery often starts with vague ideas about desired mechanical behaviors rather than precise numbers. The paper proposes MetaSymbO to bridge this by using multiple agents that translate language into geometric structures. One agent interprets the intent and sets up a starting point, another builds candidates in a special representation space, and a third provides feedback for improvements. Symbolic operators then allow targeted changes to this space to boost the structure's regularity and fit to the description. This results in designs that are more valid and novel than those from standard methods or direct use of language models.

Core claim

The central discovery is that symbolic-driven latent evolution, applied through a multi-agent system, enables the synthesis of microstructures whose geometry induces targeted behaviors as specified in free-form natural language, without explicit numerical targets, by disentangling latent factors and applying programmable operators for composition and refinement at inference time.

What carries the argument

Symbolic-driven latent evolution: programmable operators applied over disentangled latent factors to compose, modify, and refine candidate microstructures during the discovery process.

If this is right

  • Enables starting the design process with qualitative language-based intents instead of numerical property targets.
  • Yields microstructures with higher structural validity measured by symmetry and periodicity.
  • Provides better alignment with the input language guidance while keeping higher novelty in structures.
  • Validates through case studies on auxetic and high-stiffness metamaterials for practical use.

Where Pith is reading between the lines

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

  • The framework could be extended to incorporate real-time simulation feedback for even tighter property control.
  • Similar symbolic evolution techniques might apply to generating designs in related fields like photonics or acoustics.
  • Disentangling the latent space further could allow users to edit specific aspects of the design post-generation.

Load-bearing premise

The premise that a disentangled latent space combined with symbolic operators can reliably produce physically valid and intent-matching microstructures from language alone, without additional numerical optimization or filtering steps.

What would settle it

Generating a set of structures from language prompts and then performing mechanical testing or high-fidelity simulation to measure if their actual properties match the described behaviors and if the validity improvements persist without post-hoc corrections.

Figures

Figures reproduced from arXiv: 2604.27300 by Dawei Zhou, Dongqi Fu, Jianpeng Chen, Junkai Zhang, Ling Li, Wangzhi Zhan, Wei Wang, Zian Jia.

Figure 1
Figure 1. Figure 1: Truss lattice. Metamaterial Design. Metamaterials, as shown in view at source ↗
Figure 2
Figure 2. Figure 2: Overview of METASYMBO. Agent Designer translates the prompt into a scaffold, Agent Generator refines the design in latent geometric space via symbolic-driven latent evolution, and Agent Supervisor evaluates properties to provide fast iterative feedback. prompt; (2) after the Designer could successfully understand the prompt, the critical issue is how the model can translate the intent of prompt into geomet… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative analysis shows the proposed operators and latent evolution methods view at source ↗
Figure 5
Figure 5. Figure 5: Case study with METASYMBO. FE simulation denotes finite-element simulation. node/edge alignment in latent evolution. Removing Lprior or Lr also noticeably hurts validity, as both help preserve feasible structures during evolution. In contrast, Ls mainly affects the semantic perturbation and thus influences all metrics view at source ↗
Figure 4
Figure 4. Figure 4: Property prediction. Property Prediction Performance. We evaluate the introduced prop￾erty prediction model used in Supervisor view at source ↗
Figure 6
Figure 6. Figure 6: Prompt for Designer. 16 view at source ↗
Figure 7
Figure 7. Figure 7: Prompt for Supervisor. 17 view at source ↗
Figure 8
Figure 8. Figure 8: Implementation details of the disentangled encoder and decoder. After the view at source ↗
Figure 9
Figure 9. Figure 9: Samples in MetaModulus dataset. Prompts for Language Guidance. In order to evaluate the language-guidance effective￾ness, we introduce 100 design prompts to test if the model can generate effective structures that fit the prompt semantically. These prompt targets on high-level design concepts, con￾taining terms such as, “high-stiffness”, “hard material”, “extremely flexible”, etc view at source ↗
Figure 10
Figure 10. Figure 10: The prompt in Agent Supervisor for evaluation. view at source ↗
Figure 11
Figure 11. Figure 11: Visualization of two generated unit cells produced by Agent Designer. view at source ↗
Figure 12
Figure 12. Figure 12: Qualitative results of Intersection, Negation, and Mixture symbolic operators. view at source ↗
Figure 13
Figure 13. Figure 13: Prediction results of the proposed disentangled VAE on three properties. All view at source ↗
Figure 14
Figure 14. Figure 14: Second case study. 29 view at source ↗
Figure 15
Figure 15. Figure 15: Third case study. C.6 Wet Lab Results For wet-lab validation, Figures 16a and 16b present representative 3D-printed lattice spec￾imens obtained by periodically tiling the generated truss-like unit cells into macroscopic samples, consistent with the lattice-based metamaterial formulation of the framework. Figure 16a shows a small proof-of-concept print placed next to common laboratory con￾tainers, providin… view at source ↗
Figure 16
Figure 16. Figure 16: Wet-lab validation via 3D printing. Left: a proof-of-concept printed lattice with view at source ↗
read the original abstract

Metamaterial discovery seeks microstructured materials whose geometry induces targeted mechanical behavior. Existing inverse-design methods can efficiently generate candidates, but they typically require explicit numerical property targets and are less suitable for early-stage exploration, where researchers often begin with incomplete constraints and qualitative intents expressed in natural language. Large language models can interpret such intents, but they lack geometric awareness and physical property validity. To address this gap, we propose MetaSymbO, a multi-agent framework for language-guided Metamaterial discovery via Symbolic-driven latent evOlution. Specifically, MetaSymbO contains three agents: a Designer that interprets free-form design intents and retrieves a semantically consistent scaffold, a Generator that synthesizes candidate microstructures in a disentangled latent space, and a Supervisor that provides fast property-aware feedback for iterative refinement. To move beyond the limitations of reproducing known samples from literature and training data, we further introduce symbolic-driven latent evolution, which applies programmable operators over disentangled latent factors to compose, modify, and refine structures at inference time. Extensive experiments demonstrate that (i) MetaSymbO improves structural validity by up to 34% in symmetry and nearly 98% in periodicity compared to state-of-the-art baselines; (ii) MetaSymbO achieves about 6-7% higher language-guidance scores while maintaining superior structure novelty compared to advanced reasoning LLMs; (iii) qualitative analyses confirm the effectiveness of symbolic logic operators in enabling programmable semantic alignment; and (iv) realworld case studies on auxetic, high-stiffness metamaterial design further validate its practical capability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes MetaSymbO, a multi-agent framework (Designer, Generator, Supervisor) for language-guided metamaterial discovery. It uses a disentangled latent space combined with symbolic-driven latent evolution via programmable operators to generate microstructures from free-form natural language intents. The central claims are quantitative gains in structural validity (up to 34% symmetry, ~98% periodicity over baselines), 6-7% higher language-guidance scores with maintained novelty, effectiveness of symbolic operators for semantic alignment, and validation via real-world case studies on auxetic and high-stiffness metamaterials.

Significance. If the results hold and the property-aware feedback is grounded in physics, the work could meaningfully advance early-stage metamaterial exploration by allowing qualitative language inputs without requiring explicit numerical targets. The combination of multi-agent LLM reasoning with symbolic operators over disentangled latents offers a potentially generalizable way to improve validity and novelty beyond pure generative models or direct LLM generation. The case studies on auxetic and stiff designs provide a concrete test of practical utility.

major comments (3)
  1. [Abstract and Supervisor agent section] Abstract (points i and iv) and the Supervisor agent description: the 'fast property-aware feedback' is presented as enabling iterative refinement toward targeted mechanical behaviors (auxeticity, stiffness), yet no derivation, surrogate model, or simulation reference (e.g., FEM, homogenization) is given for how mechanical properties are evaluated. Geometric metrics (symmetry, periodicity) and language scores alone do not establish that the generated structures induce the intended physical responses; this is load-bearing for the central claim that language-guided symbolic evolution produces functionally valid metamaterials.
  2. [Experiments section] Experiments section (quantitative results): the reported 34% symmetry and 98% periodicity improvements require explicit baseline definitions, training data overlap checks, and statistical error analysis (e.g., standard deviations across multiple runs or seeds). Without these, it is unclear whether gains arise from the symbolic evolution or from differences in how baselines handle language inputs versus the proposed multi-agent setup.
  3. [Case studies section] Case studies on auxetic and high-stiffness designs: the qualitative validation must include quantitative mechanical property measurements (e.g., Poisson's ratio, effective modulus from simulation) rather than relying solely on visual inspection or language alignment scores. If post-hoc filtering or manual selection is applied, this would undermine the claim of end-to-end language-to-valid-structure generation.
minor comments (2)
  1. [Methods] The abstract mentions 'disentangled latent space' and 'programmable symbolic operators' without a brief equation or pseudocode example; adding a short formal description in the methods would improve accessibility.
  2. [Figures and tables] Figure captions and tables should explicitly state the number of samples, random seeds, and exact metric definitions (e.g., how periodicity is quantified) to allow reproduction.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment point by point below, clarifying our approach and indicating where the manuscript will be revised for greater rigor and clarity.

read point-by-point responses

  1. Referee: [Abstract and Supervisor agent section] Abstract (points i and iv) and the Supervisor agent description: the 'fast property-aware feedback' is presented as enabling iterative refinement toward targeted mechanical behaviors (auxeticity, stiffness), yet no derivation, surrogate model, or simulation reference (e.g., FEM, homogenization) is given for how mechanical properties are evaluated. Geometric metrics (symmetry, periodicity) and language scores alone do not establish that the generated structures induce the intended physical responses; this is load-bearing for the central claim that language-guided symbolic evolution produces functionally valid metamaterials.

    Authors: We appreciate this observation. The Supervisor provides fast feedback by combining geometric validity metrics (symmetry and periodicity, which are established proxies for mechanical isotropy and effective properties in metamaterial literature) with language alignment scores. These proxies enable iterative refinement without full physics simulation in the loop. We acknowledge that the manuscript would benefit from explicit references to how these metrics correlate with targeted behaviors such as auxeticity. We will revise the Supervisor section and abstract to include these derivations and literature citations while clarifying that full FEM is reserved for final validation to maintain computational efficiency. revision: partial

  2. Referee: [Experiments section] Experiments section (quantitative results): the reported 34% symmetry and 98% periodicity improvements require explicit baseline definitions, training data overlap checks, and statistical error analysis (e.g., standard deviations across multiple runs or seeds). Without these, it is unclear whether gains arise from the symbolic evolution or from differences in how baselines handle language inputs versus the proposed multi-agent setup.

    Authors: We agree that these details are essential for reproducibility and to isolate the contribution of symbolic latent evolution. In the revised Experiments section we will explicitly define each baseline (including their language-handling mechanisms), report training-data overlap statistics, and include standard deviations computed across multiple independent runs with different random seeds. revision: yes

  3. Referee: [Case studies section] Case studies on auxetic and high-stiffness designs: the qualitative validation must include quantitative mechanical property measurements (e.g., Poisson's ratio, effective modulus from simulation) rather than relying solely on visual inspection or language alignment scores. If post-hoc filtering or manual selection is applied, this would undermine the claim of end-to-end language-to-valid-structure generation.

    Authors: We thank the referee for highlighting this requirement. The current case studies use visual inspection and language scores to illustrate practical utility, with no post-hoc filtering or manual selection applied—all structures are direct outputs. To strengthen the claims, we will augment the case-study section with quantitative FEM-derived metrics (Poisson’s ratio for auxetic examples and effective modulus for high-stiffness examples) computed on the generated microstructures. revision: yes

Circularity Check

0 steps flagged

No significant circularity: framework claims rest on empirical validation rather than self-referential definitions or fitted predictions.

full rationale

The paper describes a multi-agent system (Designer, Generator, Supervisor) plus symbolic operators on a disentangled latent space for language-guided metamaterial generation. No equations, parameter fits, or self-citations are shown that would make the reported gains in symmetry (+34%), periodicity (+98%), or language scores (+6-7%) reduce to the method's own inputs by construction. The central claims are supported by external experimental comparisons to baselines and qualitative case studies on auxetic/high-stiffness designs; these are falsifiable outcomes, not tautologies. The derivation chain is therefore self-contained against external benchmarks, with no load-bearing self-definitional steps, fitted-input predictions, or uniqueness theorems imported from the authors' prior work.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

The central claim rests on the assumption that LLMs can reliably map free-form language to semantically consistent scaffolds and that symbolic operators over latent factors preserve physical validity; no free parameters or invented physical entities are stated.

axioms (2)
  • domain assumption Large language models can interpret free-form design intents and retrieve semantically consistent scaffolds
    Invoked for the Designer agent in the abstract.
  • domain assumption Disentangled latent space allows independent modification of geometric factors via symbolic operators
    Core premise of the Generator and symbolic evolution component.
invented entities (1)
  • Symbolic-driven latent evolution no independent evidence
    purpose: Apply programmable operators to compose and refine microstructures at inference time
    New technique introduced to move beyond reproducing training data

pith-pipeline@v0.9.0 · 5607 in / 1411 out tokens · 43697 ms · 2026-05-07T08:49:14.730843+00:00 · methodology

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

60 extracted references · 17 canonical work pages · 6 internal anchors

  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]

    Journal of Engineering Materials and Technology, 141 0 (1), January 2019

    A 149 line homogenization code for three-dimensional cellular materials written in matlab. Journal of Engineering Materials and Technology, 141 0 (1), January 2019. ISSN 0094-4289. doi:10.1115/1.4040555

    work page doi:10.1115/1.4040555 2019

  3. [3]

    How to determine composite material properties using numerical homogenization

    Erik Andreassen and Casper Schousboe Andreasen. How to determine composite material properties using numerical homogenization. Computational Materials Science, 83: 0 488--495, 2014. ISSN 0927-0256. doi:https://doi.org/10.1016/j.commatsci.2013.09.006

    work page doi:10.1016/j.commatsci.2013.09.006 2014

  4. [4]

    Mechanical properties of lattice materials via asymptotic homogenization and comparison with alternative homogenization methods

    Sajad Arabnejad and Damiano Pasini. Mechanical properties of lattice materials via asymptotic homogenization and comparison with alternative homogenization methods. International Journal of Mechanical Sciences, 77: 0 249--262, 2013. ISSN 0020-7403. doi:https://doi.org/10.1016/j.ijmecsci.2013.10.003

    work page doi:10.1016/j.ijmecsci.2013.10.003 2013

  5. [5]

    Inverting the structure--property map of truss metamaterials by deep learning

    Jan-Hendrik Bastek, Siddhant Kumar, Bastian Telgen, Rapha \"e l N Glaesener, and Dennis M Kochmann. Inverting the structure--property map of truss metamaterials by deep learning. Proceedings of the National Academy of Sciences, 119 0 (1): 0 e2111505119, 2022

    2022

  6. [6]

    Pattern recognition and machine learning, volume 4

    Christopher M Bishop and Nasser M Nasrabadi. Pattern recognition and machine learning, volume 4. Springer, 2006

    2006

  7. [7]

    Customizing graph neural networks using path reweighting

    Jianpeng Chen, Yujing Wang, Ming Zeng, Zongyi Xiang, Bitan Hou, Yunhai Tong, Ole J Mengshoel, and Yazhou Ren. Customizing graph neural networks using path reweighting. Information Sciences, 674: 0 120681, 2024 a

    2024

  8. [8]

    Yu, and Lifang He

    Jianpeng Chen, Yawen Ling, Jie Xu, Yazhou Ren, Shudong Huang, Xiaorong Pu, Zhifeng Hao, Philip S. Yu, and Lifang He. Variational graph generator for multiview graph clustering. IEEE Transactions on Neural Networks and Learning Systems, 36 0 (6): 0 11078--11091, 2025 a . doi:10.1109/TNNLS.2024.3524205

    work page doi:10.1109/tnnls.2024.3524205 2025

  9. [9]

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

    Jianpeng Chen, Wangzhi Zhan, Haohui Wang, Zian Jia, Jingru Gan, Junkai Zhang, Jingyuan Qi, Tingwei Chen, Lifu Huang, Muhao Chen, et al. Metamatbench: Integrating heterogeneous data, computational tools, and visual interface for metamaterial discovery. In Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 2, pp.\ 5334--...

    2025

  10. [10]

    Smhgc: Homophily-agnostic multi-view heterophilous graph clustering

    Jianpeng Chen, Yawen Ling, Yazhou Ren, Shufei Zhang, and Lifang He. Smhgc: Homophily-agnostic multi-view heterophilous graph clustering. Pattern Recognition, pp.\ 113668, 2026

    2026

  11. [11]

    Generative inverse design of metamaterials with functional responses by interpretable learning

    Wei Chen, Rachel Sun, Doksoo Lee, Carlos M Portela, and Wei Chen. Generative inverse design of metamaterials with functional responses by interpretable learning. Advanced Intelligent Systems, pp.\ 2400611, 2024 b

    2024

  12. [12]

    Nearest neighbor pattern classification

    Thomas Cover and Peter Hart. Nearest neighbor pattern classification. IEEE transactions on information theory, 13 0 (1): 0 21--27, 1967

    1967

  13. [13]

    Inverse design of mechanical metamaterials with target nonlinear response via a neural accelerated evolution strategy

    Bolei Deng, Ahmad Zareei, Xiaoxiao Ding, James C Weaver, Chris H Rycroft, and Katia Bertoldi. Inverse design of mechanical metamaterials with target nonlinear response via a neural accelerated evolution strategy. Advanced Materials, 34 0 (41): 0 2206238, 2022

    2022

  14. [14]

    Metamaterials: physics and engineering explorations

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

    2006

  15. [15]

    Learning with a wasserstein loss

    Charlie Frogner, Chiyuan Zhang, Hossein Mobahi, Mauricio Araya, and Tomaso Poggio. Learning with a wasserstein loss. In Advances in Neural Information Processing Systems (NeurIPS), volume 28, pp.\ 2053--2061, 2015

    2053

  16. [16]

    A latent diffusion model for protein structure generation

    Cong Fu, Keqiang Yan, Limei Wang, Wing Yee Au, Michael Curtis McThrow, Tao Komikado, Koji Maruhashi, Kanji Uchino, Xiaoning Qian, and Shuiwang Ji. A latent diffusion model for protein structure generation. In Learning on Graphs Conference, pp.\ 29--1. PMLR, 2024

    2024

  17. [17]

    Lawrence Zitnick, and Zachary Ward Ulissi

    Nate Gruver, Anuroop Sriram, Andrea Madotto, Andrew Gordon Wilson, C. Lawrence Zitnick, and Zachary Ward Ulissi. 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

  18. [18]

    Deepseek-r1 incentivizes reasoning in llms through reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1 incentivizes reasoning in llms through reinforcement learning. Nature, 645 0 (8081): 0 633--638, 2025

    2025

  19. [19]

    Rapid inverse design of metamaterials based on prescribed mechanical behavior through machine learning

    Chan Soo Ha, Desheng Yao, Zhenpeng Xu, Chenang Liu, Han Liu, Daniel Elkins, Matthew Kile, Vikram Deshpande, Zhenyu Kong, Mathieu Bauchy, et al. Rapid inverse design of metamaterials based on prescribed mechanical behavior through machine learning. Nature Communications, 14 0 (1): 0 5765, 2023

    2023

  20. [20]

    Forward and inverse design of kirigami via supervised autoencoder

    Paul Z Hanakata, Ekin D Cubuk, David K Campbell, and Harold S Park. Forward and inverse design of kirigami via supervised autoencoder. Physical Review Research, 2 0 (4): 0 042006, 2020

    2020

  21. [21]

    Gleghorn

    Thomas Hayes, Roshan Rao, Halil Akin, Nicholas J Sofroniew, Deniz Oktay, Zeming Lin, Robert Verkuil, Vincent Q Tran, Jonathan Deaton, Marius Wiggert, et al. Simulating 500 million years of evolution with a language model, july 2024. URL https://www. biorxiv. org/content/10.1101/2024.07, 1: 0 v1, 2024

    work page doi:10.1101/2024.07 2024

  22. [22]

    Geoffrey E. Hinton. Training products of experts by minimizing contrastive divergence. Neural Computation, 14 0 (8): 0 1771--1800, 08 2002. ISSN 0899-7667. doi:10.1162/089976602760128018

    work page doi:10.1162/089976602760128018 2002

  23. [23]

    GPT-4o System Card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    work page internal anchor Pith review arXiv 2024

  24. [24]

    Large language model agent as a mechanical designer, 2024

    Yayati Jadhav and Amir Barati Farimani. Large language model agent as a mechanical designer, 2024. URL https://arxiv. org/abs/2404.17525, 2024

    work page arXiv 2024

  25. [25]

    Engineering lattice metamaterials for extreme property, programmability, and multifunctionality

    Zian Jia, Fan Liu, Xihang Jiang, and Lifeng Wang. Engineering lattice metamaterials for extreme property, programmability, and multifunctionality. Journal of Applied Physics, 127 0 (15), 2020

    2020

  26. [26]

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

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. 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

  27. [27]

    Mechanical characterization and inverse design of stochastic architected metamaterials using neural operators

    Hanxun Jin, Enrui Zhang, Boyu Zhang, Sridhar Krishnaswamy, George Em Karniadakis, and Horacio D Espinosa. Mechanical characterization and inverse design of stochastic architected metamaterials using neural operators. arXiv preprint arXiv:2311.13812, 2023

    work page arXiv 2023

  28. [28]

    Identifiability of product of experts models

    Manav Kant, Eric Y Ma, Andrei Staicu, Leonard J Schulman, and Spencer Gordon. Identifiability of product of experts models. In Sanjoy Dasgupta, Stephan Mandt, and Yingzhen Li (eds.), Proceedings of The 27th International Conference on Artificial Intelligence and Statistics, volume 238 of Proceedings of Machine Learning Research, pp.\ 4492--4500. PMLR, 02-...

    2024

  29. [29]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2014

    2014

  30. [30]

    Data-driven design for metamaterials and multiscale systems: a review

    Doksoo Lee, Wei Chen, Liwei Wang, Yu-Chin Chan, and Wei Chen. Data-driven design for metamaterials and multiscale systems: a review. Advanced Materials, 36 0 (8): 0 2305254, 2024

    2024

  31. [31]

    Geometry informed tokenization of molecules for language model generation, 2025 a

    Xiner Li, Limei Wang, Youzhi Luo, Carl Edwards, Shurui Gui, Yuchao Lin, Heng Ji, and Shuiwang Ji. Geometry informed tokenization of molecules for language model generation, 2025 a . URL https://openreview.net/forum?id=HbZrxBXzks

    2025

  32. [32]

    Large language-geometry model: When LLM meets equivariance

    Zongzhao Li, Jiacheng Cen, Bing Su, Tingyang Xu, Yu Rong, Deli Zhao, and Wenbing Huang. Large language-geometry model: When LLM meets equivariance. In Forty-second International Conference on Machine Learning, 2025 b . URL https://openreview.net/forum?id=Qi9rPcdiTY

    2025

  33. [33]

    Equiformer: Equivariant graph attention transformer for 3d atomistic graphs

    Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. In ICLR, 2023

    2023

  34. [34]

    DeepSeek-V3 Technical Report

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report. arXiv preprint arXiv:2412.19437, 2024

    work page internal anchor Pith review arXiv 2024

  35. [35]

    Spherical message passing for 3d molecular graphs

    Yi Liu, Limei Wang, Meng Liu, Yuchao Lin, Xuan Zhang, Bora Oztekin, and Shuiwang Ji. Spherical message passing for 3d molecular graphs. In ICLR, 2022

    2022

  36. [36]

    Exploring the property space of periodic cellular structures based on crystal networks

    Thomas S Lumpe and Tino Stankovic. 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

  37. [37]

    Deep learning generative model for crystal structure prediction

    Xiaoshan Luo, Zhenyu Wang, Pengyue Gao, Jian Lv, Yanchao Wang, Changfeng Chen, and Yanming Ma. Deep learning generative model for crystal structure prediction. npj Computational Materials, 10 0 (1): 0 254, 2024 a

    2024

  38. [38]

    Towards symmetry-aware generation of periodic materials

    Youzhi Luo, Chengkai Liu, and Shuiwang Ji. Towards symmetry-aware generation of periodic materials. NeurIPS, 36, 2024 b

    2024

  39. [39]

    Probabilistic representation and inverse design of metamaterials based on a deep generative model with semi-supervised learning strategy

    Wei Ma, Feng Cheng, Yihao Xu, Qinlong Wen, and Yongmin Liu. Probabilistic representation and inverse design of metamaterials based on a deep generative model with semi-supervised learning strategy. Advanced Materials, 31 0 (35): 0 1901111, 2019

    2019

  40. [40]

    Braza, Ryan-Rhys Griffiths, Manvitha Ponnapati, Albert Bou, Jon M Laurent, Ori Kabeli, Geemi Wellawatte, Sam Cox, Samuel G Rodriques, and Andrew White

    Siddharth Narayanan, James D. Braza, Ryan-Rhys Griffiths, Manvitha Ponnapati, Albert Bou, Jon M Laurent, Ori Kabeli, Geemi Wellawatte, Sam Cox, Samuel G Rodriques, and Andrew White. Aviary: Training language agents on challenging scientific tasks. In Scaling Self-Improving Foundation Models without Human Supervision, 2025. URL https://openreview.net/forum...

    2025

  41. [41]

    Deep learning for size-agnostic inverse design of random-network 3d printed mechanical metamaterials

    Helda Pahlavani, Kostas Tsifoutis-Kazolis, Mauricio C Saldivar, Prerak Mody, Jie Zhou, Mohammad J Mirzaali, and Amir A Zadpoor. Deep learning for size-agnostic inverse design of random-network 3d printed mechanical metamaterials. Advanced Materials, 36 0 (6): 0 2303481, 2024

    2024

  42. [42]

    Optical metamaterials: fundamentals and applications, 2010

    Dilip D Paul. Optical metamaterials: fundamentals and applications, 2010

    2010

  43. [43]

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis

    Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas M \"u ller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023

    work page internal anchor Pith review arXiv 2023

  44. [44]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  45. [45]

    Inverse design of discrete mechanical metamaterials

    Henrik Ronellenfitsch, Norbert Stoop, Josephine Yu, Aden Forrow, and J \"o rn Dunkel. Inverse design of discrete mechanical metamaterials. Physical Review Materials, 3 0 (9): 0 095201, 2019

    2019

  46. [46]

    Gemini: A Family of Highly Capable Multimodal Models

    Gemini Team. Gemini: A family of highly capable multimodal models, 2025. URL https://arxiv.org/abs/2312.11805

    work page internal anchor Pith review arXiv 2025

  47. [47]

    Machine learning-based prediction and inverse design of 2d metamaterial structures with tunable deformation-dependent poisson's ratio

    Jie Tian, Keke Tang, Xianyan Chen, and Xianqiao Wang. Machine learning-based prediction and inverse design of 2d metamaterial structures with tunable deformation-dependent poisson's ratio. Nanoscale, 14 0 (35): 0 12677--12691, 2022

    2022

  48. [48]

    A multi-agent framework integrating large language models and generative ai for accelerated metamaterial design

    Jie Tian, Martin Taylor Sobczak, Dhanush Patil, Jixin Hou, Lin Pang, Arunachalam Ramanathan, Libin Yang, Xianyan Chen, Yuval Golan, Xiaoming Zhai, et al. A multi-agent framework integrating large language models and generative ai for accelerated metamaterial design. arXiv preprint arXiv:2503.19889, 2025

    work page arXiv 2025

  49. [49]

    LLaMA: Open and Efficient Foundation Language Models

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023

    work page internal anchor Pith review arXiv 2023

  50. [50]

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

    Yusong Wang, Tong Wang, Shaoning Li, Xinheng He, Mingyu Li, Zun Wang, Nanning Zheng, Bin Shao, and Tie-Yan Liu. Enhancing geometric representations for molecules with equivariant vector-scalar interactive message passing. Nature Communications, 15 0 (1): 0 313, 2024

    2024

  51. [51]

    Jaakkola

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi S. Jaakkola. Crystal diffusion variational autoencoder for periodic material generation. In ICLR, 2022. URL https://openreview.net/forum?id=03RLpj-tc_

    2022

  52. [52]

    Qwen2.5 Technical Report

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Tingyu X...

    work page internal anchor Pith review arXiv 2024

  53. [53]

    Data-efficient and interpretable inverse materials design using a disentangled variational autoencoder

    Cheng Zeng, Zulqarnain Khan, and Nathan L Post. Data-efficient and interpretable inverse materials design using a disentangled variational autoencoder. AI Mater., 2025. doi:https://doi.org/10.55092/aimat20250002

    work page doi:10.55092/aimat20250002 2025

  54. [54]

    Unimate: A unified model for mechanical metamaterial generation, property prediction, and condition confirmation

    Wangzhi Zhan, Jianpeng Chen, Dongqi Fu, and Dawei Zhou. Unimate: A unified model for mechanical metamaterial generation, property prediction, and condition confirmation. In Forty-second International Conference on Machine Learning, 2025. URL https://openreview.net/forum?id=X7uVxeFS9A

    2025

  55. [55]

    Ritchie, and Jian Xiong

    Jian Zhao, Robert O. Ritchie, and Jian Xiong. High-performance programmable combinatorial lattice materials. Journal of the Mechanics and Physics of Solids, 206: 0 106351, 2025. ISSN 0022-5096. doi:https://doi.org/10.1016/j.jmps.2025.106351. URL https://www.sciencedirect.com/science/article/pii/S0022509625003254

    work page doi:10.1016/j.jmps.2025.106351 2025

  56. [56]

    Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling

    Li Zheng, Konstantinos Karapiperis, Siddhant Kumar, and Dennis M Kochmann. Unifying the design space and optimizing linear and nonlinear truss metamaterials by generative modeling. Nature Communications, 14 0 (1): 0 7563, 2023

    2023

  57. [57]

    Bindgpt: A scalable framework for 3d molecular design via language modeling and reinforcement learning

    Artem Zholus, Maksim Kuznetsov, Roman Schutski, Rim Shayakhmetov, Daniil Polykovskiy, Sarath Chandar, and Alex Zhavoronkov. Bindgpt: A scalable framework for 3d molecular design via language modeling and reinforcement learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pp.\ 26083--26091, 2025

    2025

  58. [58]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should not add it explicitly Type <Return> for now, but then later remove the command n...

  59. [59]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@first@sw \@firstoftwo \@ifundefined NAT@b*@#2 \@firstoftwo @num @NAT@ctr \@secondoft...

  60. [60]

    EԥK QAAAOk||ر9 [cƌW_ ew 7# x gC'OVٲe-4`fy. 3f- ]Hk׮Չ ' k޼yh 4hӱ ]f ,X`>b լY-sn-߿||| 3f85V\\. fc!G_U#G @0x`jʪ:

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibsetup #1 @NAT@ctr @ @openbib .11em \@plus.33em \@minus.07em 4000 4000 `\.\@m @bibit...