Toward a Robust and Generalizable Metamaterial Foundation Model

Dongseok Lee; Dosung Lee; Jongbin Yu; Namjung Kim; Sung Woong Cho; Yesol Park; Youngjoon Hong

REVIEW 5 major objections 5 minor 42 references

Toward a Robust and Generalizable Metamaterial Foundation Model

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper introduces MetaFO, a transformer trained once on 10,230 simulated stress-strain curves of mechanical metamaterials, and claims it predicts unseen structures, fills missing and noisy data, and designs unit cells from target…

desk verdict MetaFO is a plausible interpolator within one combinatorial family, but the foundation-model and OOD claims are untested because held-out points stay inside the training hull. read the letter →

arxiv 2507.02436 v1 pith:ELZYQ6DN submitted 2025-07-03 cs.CE cs.AIphysics.optics

Namjung Kim , Dongseok Lee , Jongbin Yu , Sung Woong Cho , Dosung Lee , Yesol Park , Youngjoon Hong This is my paper

classification cs.CEcs.AIphysics.optics MSC 68T0774B2074S05

keywords foundationmodelin-contextlearningmechanicalmetamaterialsinversedesignBayesiantransformerneuraloperatorstress-strainpredictionzero-shotgeneralization

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 introduces MetaFO, a transformer model trained once on 10,230 simulated stress-strain curves of mechanical metamaterials. The central claim is that a single pretrained model learns a general operator that maps any base material's stress-strain behavior to the structural response of any unit-cell geometry drawn from ten basis cells, enabling zero-shot prediction for unseen combinations. The same model is claimed to reconstruct missing stretches of response curves, suppress Gaussian measurement noise, and invert the map to generate unit-cell connectivity graphs from a target mechanical response. If the claims hold, task-specific retraining and separate forward/inverse models become unnecessary, and metamaterial design can be prompt-based within one architecture.

What carries the argument

The central object is the operator that maps a material property function (stress as a function of strain for a base material) to a unit-cell response function (the stress-strain curve of the metamaterial). MetaFO approximates this operator with a transformer encoder-decoder: dual multilayer-perceptron encoders embed discretized strain-stress tokens, a global context vector is formed by averaging solution embeddings of the known prompt pairs, and a learnable noise bank generates dynamic noise embeddings conditioned on that context. Cross-attention between decoder queries and encoder memory implements in-context learning, so predictions adapt to each prompt without weight updates. This operator formulation is what lets one model carry forward prediction, missing-data completion, denoising, and inverse design.

What would settle it

Compression-test 3D-printed samples of held-out unit cells made from the same commercial elastomer classes used in training, compare measured stress-strain curves to MetaFO's zero-shot predictions, and check the yield drop, plateau, and densification branches; errors well above the reported 2-4% range would falsify the generalization claim.

Watch

Extended reading notes

Core claim

MetaFO's central discovery is that a single neural operator, pretrained on finite-element data, can generalize across metamaterial geometries and base materials without retraining, and the same learned representation supports forward prediction, data completion, denoising, and inverse design in one framework. Treating each unit cell as an operator that maps material property functions to structural response functions, MetaFO uses prompt pairs of known material-response data as context; attention over this context lets it predict responses for new materials or geometries, even outside the training range. The paper reports zero-shot prediction errors below 2.5% for unseen unit-cell combinations, interpolation and extrapolation errors below 2%, prediction errors below 4% with up to 10% injected Gaussian noise, and inverse design that recovers unit-cell node connectivity from target stress-strain pairs.

Load-bearing premise

Every reported accuracy number assumes the finite-element stress-strain curves used as labels faithfully represent real compression behavior of 3D-printed elastomer metamaterials in the elastic, yielding, plateau, and densification regimes; if those synthetic labels are biased, the zero-shot and inverse-design results inherit the bias.

Editorial extensions

If this is right

Zero-shot stress-strain prediction for unseen unit cells and base materials becomes possible after a single pretraining, removing the need for task-specific retraining.
The same model can complete missing segments of response curves, including extrapolation beyond the training strain range, which is useful for experimental data with dropouts.
Noise-tolerant inference means lower-precision measurement setups could feed the model directly, with prediction error staying below injected noise levels.
Inverse design from target mechanical responses produces unit-cell connectivity graphs, unifying forward evaluation and generative design in one architecture.
Prompt-based interaction with a mechanics model, analogous to language-model prompting, could speed up simulation-driven design workflows.

Reading between the lines

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

The operator formulation is not restricted to compression or elasticity; the same prompt-based transformer could in principle learn thermal, acoustic, or electromagnetic structure-property maps if trained on corresponding simulation data, though the paper does not test this.
The reported accuracy gain from adding 20-40% noisy training samples suggests the noise-injection mechanism acts as a regularizer; a direct test would be to ablate the dynamic noise bank and compare generalization on clean data.
Inverse design currently outputs binary graphs by thresholding a continuous connectivity matrix; extending it to continuous geometry, graded strut thickness, or full stress fields is a natural next step not demonstrated here.
Because the pretraining set is small by foundation-model standards, scaling to more basis cells, more materials, and richer deformation modes is the obvious test of whether the observed generalization is a scaling effect or specific to this dataset.

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.

Desk Editor's Note

read the letter

Short version: the actual contribution is a prompt-based transformer that predicts stress-strain curves for unit cells composed of ten hand-picked basis cells, and it does plausibly well at interpolation within that family. The 'foundation model' and 'zero-shot OOD' framing outruns the evidence. The architecture is clearly specified (dual encoders, transformer, dynamic noise bank, inverse head that outputs adjacency), the dataset is a nontrivial 10,230 FEM samples, and the four tasks—forward, missing data, noise, inverse design—are unified in one framework. That is a useful engineering contribution, and the plots suggest the model genuinely learns to combine basis cells and materials.

The soft spots are substantive. First, the generalization test is not what the abstract promises. Held-out unit cells are subsets of the same ten basis cells, and held-out materials are interpolations of three Formlabs elastomers. The model never sees a topology or material class outside its training hull, so the results demonstrate interpolation, not OOD generalization. Second, the paper claims 'elastic and plastic regimes' but the SI explicitly says the simulations use a hyperelastic Yeoh model (because plasticity models had convergence issues). There is no plasticity in the labels, so the language about yielding and post-yield plateau is not backed by a plastic constitutive model. That is a mismatch between claims and evidence. Third, there is no code or data release; 'available upon request' is insufficient for a numerical methods paper. Fourth, the inverse design results are entirely qualitative—no quantitative connectivity accuracy. The baselines in Fig. 2c are not named, so the '10–100x' claim is unverifiable.

The paper deserves a serious referee, but it needs major revision. The narrow interpolation claim is plausible and could be publishable after the authors release artifacts, name baselines, add a genuine OOD test (e.g., hold out one basis cell entirely, not just subsets), remove or substantiate the plasticity claim, and quantify inverse design. I would not cite it in its current form, and I would not bring it to reading group as a model of how to validate a foundation model.

Referee Report

5 major / 5 minor

Summary. The manuscript introduces MetaFO, a transformer-based encoder-decoder trained on 10,230 synthetic stress-strain curves obtained from ABAQUS simulations of 1,023 unit cells, formed as all non-empty combinations of 10 hand-picked basis cells, combined with 10 materials (three commercial elastomers and seven interpolated virtual materials). The model is prompted with material property/unit-cell response pairs and predicts response curves for new configurations, reconstructs missing or noisy data, and generates connectivity graphs for inverse design. The paper claims zero-shot and out-of-distribution generalization, an operator-based foundation model for mechanics, and an implicit Bayesian posterior predictive interpretation of the trained network.

Significance. If the broad claims were supported, the work would be a useful step toward unified, prompt-based forward and inverse design of mechanical metamaterials. The architecture and task framing are coherent, and the authors are explicit about their dataset construction and simulation pipeline. However, the paper does not ship code or data, and the central generalization claim is tested only inside the combinatorial/material hull used for training, so the 'foundation model' and 'OOD' conclusions are not presently established. The contributions that are likely robust — accurate in-family interpolation for this specific dataset and a unified architecture for four tasks — are still worth reporting after substantial reframing and added evaluation.

major comments (5)

[SI S6 and Figs. 2-4] The zero-shot/OOD claims are not tested by the current split. SI S6 partitions the unit-cell indices and material indices 80:20, but the held-out unit cells are subsets of the same 10 basis cells whose individual responses appear throughout training, and the held-out materials are interpolations of the same three commercial elastomers described in SI S2. An 'unseen' configuration is therefore an unseen subset, not an unseen topology or material class, and the reported accuracies are consistent with the model learning a combinatorial/material interpolation rule. The abstract's claims of 'diverse, unseen combinations' and 'OOD conditions' should either be backed by a truly held-out basis-cell set or an independent metamaterial family, or the claims must be reframed as interpolation within the training family.
[Methods (Theoretical basis) and SI S2-S3] The manuscript repeatedly claims behavior in both 'elastic and plastic regimes', including 'abrupt stress drop after yielding' (Fig. 2d), but SI S2 explicitly states that hyperelastic Yeoh models were chosen because 'plasticity models often suffer from convergence issues', and SI S3 describes Timoshenko beam elements with a hyperelastic material. No plasticity model is used. The post-yield plateau and densification are therefore geometric/buckling phenomena in a hyperelastic simulation, not material plasticity. The claims about plastic regimes should be corrected, and either a plastic material model should be added or the wording changed to 'post-buckling/nonlinear elastic response'.
[Fig. 5 and Task 4] The inverse-design evaluation is qualitative. Fig. 5b shows predicted and ground-truth graphs but provides no quantitative metric (e.g., edge accuracy, F1, graph edit distance), no sensitivity analysis for the thresholding step, and no comparison with baselines. Since 'excels in nonlinear inverse design' is a central advertised capability, the authors should add quantitative evaluation over many test cases and report how performance varies with the adjacency threshold.
[Code availability and Methods] Reproducibility is a load-bearing issue for a machine-learning paper. The data and code are only 'available from the corresponding author upon request', and the Methods do not give training hyperparameters, optimizer settings, number of epochs, seeds, or hardware. The quantitative claims in Figs. 2-4 cannot be checked or reproduced. A public repository with data generation scripts, model code, and evaluation scripts is needed, or at minimum a detailed training configuration in the main text or SI.
[Methods (Theoretical basis)] The Bayesian interpretation is asserted rather than demonstrated. The text states that minimizing MSE is equivalent to minimizing KL divergence under Gaussian noise and that the transformer approximates a posterior predictive operator within 'an arbitrarily small epsilon', but no proof, convergence argument, or empirical calibration is provided. The paper advertises 'probabilistic' predictions, yet no predictive intervals or uncertainty-calibration plots are shown. Either provide calibration experiments or soften the claims to describe the model as trained with MSE and simply inspired by Bayesian inference.

minor comments (5)

[SI S1] The SI contains repeated typos in the geometry section, including 'yiedls', 'desing', 'gnerating', 'deisng', 'convectional', and 'unitcell'; these should be corrected.
[Algorithm 1] Two different algorithms are both labeled 'Algorithm 1' (the architecture flow and the training procedure); renumber the training procedure as Algorithm 2.
[Fig. 4d-e] The caption in Fig. 4e calls the accuracy improvement a 'Paradoxical performance gain', while the main text says 'Surprisingly'. Adding noisy data can act as a regularizer; the authors should explain this effect rather than present it as unexplained.
[SI S4] The caption mentions experimental compression tests with DIC and a 2x2x2 supercell, but no corresponding experimental results are shown. Clarify whether those data are included elsewhere or remove the reference to them.
[Methods (Operator learning framework)] Notation is inconsistent: the main text uses f_m, f_{u,m}, sigma, epsilon, and T interchangeably for functions and their discretizations. Please define all symbols once in one place before first use.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central results are held-out FEM predictions, and the Bayesian framing is an interpretation rather than a fitted input.

full rationale

The paper's derivation chain is self-contained with respect to circularity. MetaFO is pretrained once on FEM-generated material/response pairs, and all four tasks are evaluated against held-out unit cells and materials drawn from the same combinatorial and material hull. Because the held-out responses were never used in training, the reported accuracies are genuine empirical predictions rather than quantities recovered from the model's own fitted parameters. The Bayesian/posterior-predictive section is an interpretive overlay: the paper correctly notes that MSE minimization is equivalent to KL-divergence minimization under Gaussian observational noise, and then describes the trained transformer as an approximate posterior predictive. This is a standard equivalence, and it is not used as an input to any fit or as a premise that forces the reported results. The self-citations (refs. 39-41) are contextual and do not carry the load of the generalization or inverse-design claims. The main scientific weakness is that the test protocol stays inside the generative support: 'unseen' unit cells are non-empty combinations of the same ten basis cells, and the seven 'virtual' materials are interpolations of three commercial elastomers. That is a generalization-testing limitation, not a circularity, because the model is not constructed to output the held-out answer by definition. No equation reduces to another by construction, no fitted parameter is renamed as a prediction, and no uniqueness or ansatz is imported from the authors' prior work. The appropriate verdict is therefore no significant circularity.

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

The central claims rest on simulated labels generated by one hyperelastic FEM pipeline, a hand-picked basis set of 10 unit cells, and an interpolated family of 3 commercial elastomers. No free physical constants are fit, but the design space and material space are chosen by the authors, and the Bayesian interpretation is asserted rather than derived. These choices bound the claims to a closed synthetic family unless independently validated.

free parameters (4)

Basis unit cell set = 10 hand-selected cubic-symmetry cells (Fig. S1)
The entire topology space is the power set of these 10 cells; test 'unseen' cells are linear combinations of the same basis, so all generalization claims are conditional on this hand-picked set.
Virtual material interpolation = 7 Ex materials interpolated from 3 Formlabs elastomers (Fig. S2)
Of the 10 base materials, 7 are not independent; 'unseen' materials lie on low-dimensional interpolation paths, not in a truly diverse material space.
Strain sampling grid = 21 equally spaced points per stress-strain curve
All functions are represented as 21-vectors; the operator learning claim is for this fixed discretization and may change with finer or coarser sampling.
Adjacency threshold for inverse design = Not reported
Continuous connectivity scores in matrix A are converted to binary edges by an unspecified threshold; the quality of recovered graphs depends on this choice.

assumptions (4)

domain assumption Abaqus/Explicit FEM with Timoshenko beam elements and hyperelastic Yeoh material gives ground-truth stress-strain curves for metamaterial unit cells.
All training and evaluation labels come from this simulation pipeline; no independent experimental validation is shown for the 10,230 curves beyond four illustrative printed specimens in SI S4.
domain assumption A hyperelastic model captures the post-yield plateau, densification, and plastic regimes claimed in the paper.
SI S2 uses hyperelasticity to avoid plasticity convergence problems, so the claimed plastic-regime coverage rests on an assumption the text simultaneously disclaims.
ad hoc to paper MSE training plus prompt-conditioned transformer attention approximates Bayesian posterior predictive inference over operators.
The theoretical section asserts the transformer is an approximate MAP and posterior predictive estimator, but no posterior sampling or uncertainty calibration is performed; this is an interpretive premise, not a proven property.
domain assumption All non-empty linear combinations of the 10 basis unit cells form a representative design space for mechanical metamaterials.
The 1,023 topologies and the generalization claims are entirely defined by this basis set; there is no evidence that it spans diverse graph geometries outside the family.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Toward a Robust and Generalizable Metamaterial Foundation Model." pith.science (2026). https://pith.science/paper/ELZYQ6DN

@misc{pith2026250702436,
  author       = {Pith},
  title        = {Pith review of: Toward a Robust and Generalizable Metamaterial Foundation Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ELZYQ6DN}},
  note         = {Machine review of arXiv:2507.02436}
}

read the original abstract

Advances in material functionalities drive innovations across various fields, where metamaterials-defined by structure rather than composition-are leading the way. Despite the rise of artificial intelligence (AI)-driven design strategies, their impact is limited by task-specific retraining, poor out-of-distribution(OOD) generalization, and the need for separate models for forward and inverse design. To address these limitations, we introduce the Metamaterial Foundation Model (MetaFO), a Bayesian transformer-based foundation model inspired by large language models. MetaFO learns the underlying mechanics of metamaterials, enabling probabilistic, zero-shot predictions across diverse, unseen combinations of material properties and structural responses. It also excels in nonlinear inverse design, even under OOD conditions. By treating metamaterials as an operator that maps material properties to structural responses, MetaFO uncovers intricate structure-property relationships and significantly expands the design space. This scalable and generalizable framework marks a paradigm shift in AI-driven metamaterial discovery, paving the way for next-generation innovations.

Figures

Figures reproduced from arXiv: 2507.02436 by the authors.

**Figure 3.** Missing data prediction performance of MetaFO. [PITH_FULL_IMAGE:figures/full_fig_p020_3.png] view at source ↗

**Figure 4.** Robustness of MetaFO to data uncertainty and measurement noise. (a) Noise injection settings: (i) within the database and (ii) within prompt inputs, affecting both material and structural responses. (b) Predicted compressive behavior under 5% Gaussian noise shows strong agreement with ground truth across four unit cell types. (c) Prediction accuracy under increasing prompt noise (3–10%) and prompt counts (3–5); Meta… view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: Inverse design of metamaterial unit cells using MetaFO. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png]

**Figure 6.** Figure 6: Operator learning via discretized function mapping in MetaFO. [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 37 canonical work pages

[1]

Theoretical basis of MetaFO In this work, we also address the setting in which the dataset includes observational noise. Under such conditions, the discrepancy between the predicted and true response functions for a given prompt can be more appropriately quantified using the Kullback–Leibler (KL) divergence rather than a deterministic mean squared error (...

work page 2023
[2]

From metamaterials to metadevices

Zheludev NI, Kivshar YS. From metamaterials to metadevices. Nat Mater. 11, 917–924 (2012)

work page 2012
[3]

& Kellett, A

O’Halloran, S., Pandit, A., Heise, A. & Kellett, A. Two-photon polymerization: Fundamentals, materials, and chemical modification strategies. Adv. Sci. 10, 2204072 (2023)

work page 2023
[4]

Cumpston, B. H. et al. Two-photon polymerization initiators for three-dimensional optical data storage and microfabrication. Nature 398, 51–54 (1999)

work page 1999
[5]

Vyatskikh, A., Delalande, S., Kudo, A., Zhang, X., Portela, C. M. & Greer, J. R. Additive manufacturing of 3D nano-architected metals. Nat. Commun. 9, 593 (2018)

work page 2018
[6]

Cui, H. et al. Design and printing of proprioceptive three-dimensional architected robotic metamaterials. Science 376, 1287–1293 (2022)

work page 2022
[7]

& Raney, J

He, Q., Ferracin, S. & Raney, J. R. Programmable responsive metamaterials for mechanical computing and robotics. Nat. Comput. Sci. 4, 567–573 (2024)

work page 2024
[8]

& Ameer, G

Wang, H., Huddleston, S., Yang, J. & Ameer, G. A. Enabling proregenerative medical devices via citrate-based biomaterials: Transitioning from inert to regenerative biomaterials. Adv. Mater. 36, 2306326 (2024)

work page 2024

Show all 42 references

[9]

X., Wang, X

Xiong, H., Deng, J., Yang, Q., Wang, B. X., Wang, X. & Zhang, H. A metamaterial energy power detector based on electromagnetic energy harvesting technology. ACS Appl. Electron. Mater. 6, 1204–1210 (2024)

2024
[10]

I., Karavasili, C

Gioumouxouzis, C. I., Karavasili, C. & Fatouros, D. G. Recent advances in pharmaceutical dosage forms and devices using additive manufacturing technologies. Drug Discov. Today 24, 636–643 (2019)

2019
[11]

& Wen, J

Sheng, P., Fang, X., Yu, D. & Wen, J. Mitigating aeroelastic vibration of strongly nonlinear metamaterial supersonic wings under high temperature. Nonlinear Dyn. (2024). doi:10.1007/s11071-024-10278-7

2024 doi
[12]

& Palma, G

Iemma, U. & Palma, G. Design of metacontinua in the aeroacoustic spacetime. Sci. Rep. 10, 17926 (2020)

2020
[13]

Hess, O. et al. Active nanoplasmonic metamaterials. Nat. Mater. 11, 573–584 (2012)

2012
[14]

Padilla, W. J. & Averitt, R. D. Imaging with metamaterials. Nat. Rev. Phys. 4, 85–100 (2022)

2022
[15]

Shaikeea, A. J. D., Cui, H., O’Masta, M., Zheng, X. R. & Deshpande, V . S. The toughness of mechanical metamaterials. Nat. Mater. 21, 297–304 (2022)

2022
[16]

& Gumbsch, P

Fang, X., Wen, J., Cheng, L., Yu, D., Zhang, H. & Gumbsch, P. Programmable gear-based mechanical metamaterials. Nat. Mater. 21, 869–876 (2022)

2022
[17]

& Gao, L

Wang, Y., Sha, W., Xiao, M., Qiu, C. & Gao, L. Deep-learning-enabled intelligent design of thermal metamaterials. Adv. Mater. 35, 2302387 (2023)

2023
[18]

Sha, W. et al. Robustly printable freeform thermal metamaterials. Nat. Commun. 12, 7236 (2021)

2021
[19]

Jin, P. et al. Tunable liquid–solid hybrid thermal metamaterials with a topology transition. Proc. Natl. Acad. Sci. U.S.A. 120, e2217068120 (2023)

2023
[20]

Xu, J. et al. Acoustic metamaterials-driven transdermal drug delivery for rapid and on-demand 16 management of acute disease. Nat. Commun. 14, 655 (2023)

2023
[21]

Wu, L. et al. Modular design for acoustic metamaterials: Low-frequency noise attenuation. Adv. Funct. Mater. 32, 2105712 (2022)

2022
[22]

& Stucker, B

Gibson, I., Rosen, D. & Stucker, B. Design for additive manufacturing. in Additive Manufacturing Technologies 555–607 (Springer, 2021)

2021
[23]

& Tasoglu, S

Tezsezen, E., Yigci, D., Ahmadpour, A. & Tasoglu, S. AI-based metamaterial design. ACS Appl. Mater. Interfaces 16, 29547–29569 (2024)

2024
[24]

C., Zhou, J., Mirzaali, M

Pahlavani, H., Amani, M., Saldívar, M. C., Zhou, J., Mirzaali, M. J. & Zadpoor, A. A. Deep learning for the rare-event rational design of 3D printed multi-material mechanical metamaterials. Commun. Mater. 3, 48 (2022)

2022
[25]

& Wang, P

Zeng, Q., Zhao, Z., Lei, H. & Wang, P. A deep learning approach for inverse design of gradient mechanical metamaterials. Int. J. Mech. Sci. 240, 107920 (2023)

2023
[26]

& Mukhopadhyay, T

Sinha, P. & Mukhopadhyay, T. Programmable multi-physical mechanics of mechanical metamaterials. Mater. Sci. Eng. R Rep. 155, 100745 (2023)

2023
[27]

Huang, K. et al. A foundation model for clinician-centered drug repurposing. Nat. Med. 30, 3601–3613 (2024)

2024
[28]

Chen, R. J. et al. Towards a general-purpose foundation model for computational pathology. Nat. Med. 30, 850–862 (2024)

2024
[29]

Chang, J. & Ye, J. C. Bidirectional generation of structure and properties through a single molecular foundation model. Nat. Commun. 15, 1904 (2024)

2024
[30]

Herde, M. et al. Poseidon: Efficient foundation models for PDES. Adv. Neural Inf. Process. Syst. 36, 72525–72624 (2024)

2024
[31]

Jin, H. et al. Mechanical characterization and inverse design of stochastic architected metamaterials using neural operators. Adv. Mater. 2420063 (2024). doi:10.1002/adma.202420063

2024 doi
[32]

Azizzadenesheli, K. et al. Neural operators for accelerating scientific simulations and design. Preprint at https://arxiv.org/abs/2309.15325 (2024)

2024 arXiv
[33]

& Karniadakis, G

Lu, L., Jin, P. & Karniadakis, G. E. DeepONet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. Nat. Mach. Intell. 3, 218–229 (2021)

2021
[34]

Brown, T. B. et al. Language models are few-shot learners. Preprint at https://arxiv.org/abs/2005.14165 (2020)

2020 arXiv
[35]

Wei, J. et al. Chain-of-thought prompting elicits reasoning in large language models. Preprint at https://arxiv.org/abs/2201.11903 (2022)

2022 arXiv
[36]

& Valiant, G

Garg, S., Tsipras, D., Liang, P. & Valiant, G. What can transformers learn in-context? A case study of simple function classes. Adv. Neural Inf. Process. Syst. 35, 30472–30488 (2022)

2022
[37]

Radford, A. et al. Learning transferable visual models from natural language supervision. in Proc. 38th 17 Int. Conf. on Machine Learning 8748–8763 (PMLR, 2021)

2021
[38]

Sanh, V . et al. Multitask prompted training enables zero-shot task generalization. Preprint at https://arxiv.org/abs/2110.08207 (2021)

2021 arXiv
[39]

G., Hollmann, N., Berlin, C., Arango, S

Müller, S. G., Hollmann, N., Berlin, C., Arango, S. P., Grabocka, J. & Hutter, F. Meta-learning to perform Bayesian inference in a single forward propagation. Adv. Neural Inf. Process. Syst. 34, 23838–23850 (2021)

2021
[40]

Kang, M. et al. MaD-scientist: AI-based scientist solving convection-diffusion-reaction equations using massive PINN-based prior data. Preprint at https://arxiv.org/abs/2410.06442 (2024)

2024 arXiv
[41]

& Hong, Y

Kim, N., Lee, D. & Hong, Y . Data-efficient deep generative model with discrete latent representation for high-fidelity digital materials. ACS Mater. Lett. 5, 730–737 (2023)

2023
[42]

& Hong, Y

Kim, N., Lee, D., Kim, C., Lee, D. & Hong, Y . Simple arithmetic operation in latent space can generate a novel three-dimensional graph metamaterials. NPJ Comput. Mater. 10, 236 (2024). 18 Figures Figure

2024

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Theoretical basis of MetaFO In this work, we also address the setting in which the dataset includes observational noise. Under such conditions, the discrepancy between the predicted and true response functions for a given prompt can be more appropriately quantified using the Kullback–Leibler (KL) divergence rather than a deterministic mean squared error (...

work page 2023

[2] [2]

From metamaterials to metadevices

Zheludev NI, Kivshar YS. From metamaterials to metadevices. Nat Mater. 11, 917–924 (2012)

work page 2012

[3] [3]

& Kellett, A

O’Halloran, S., Pandit, A., Heise, A. & Kellett, A. Two-photon polymerization: Fundamentals, materials, and chemical modification strategies. Adv. Sci. 10, 2204072 (2023)

work page 2023

[4] [4]

Cumpston, B. H. et al. Two-photon polymerization initiators for three-dimensional optical data storage and microfabrication. Nature 398, 51–54 (1999)

work page 1999

[5] [5]

Vyatskikh, A., Delalande, S., Kudo, A., Zhang, X., Portela, C. M. & Greer, J. R. Additive manufacturing of 3D nano-architected metals. Nat. Commun. 9, 593 (2018)

work page 2018

[6] [6]

Cui, H. et al. Design and printing of proprioceptive three-dimensional architected robotic metamaterials. Science 376, 1287–1293 (2022)

work page 2022

[7] [7]

& Raney, J

He, Q., Ferracin, S. & Raney, J. R. Programmable responsive metamaterials for mechanical computing and robotics. Nat. Comput. Sci. 4, 567–573 (2024)

work page 2024

[8] [8]

& Ameer, G

Wang, H., Huddleston, S., Yang, J. & Ameer, G. A. Enabling proregenerative medical devices via citrate-based biomaterials: Transitioning from inert to regenerative biomaterials. Adv. Mater. 36, 2306326 (2024)

work page 2024