acceptodds
Withdrawn from ICLR 2027

TrussOrbit: Edge-Orbit Sequence Modeling for Truss Metamaterial Inverse Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inverse design of periodic truss metamaterials requires accurate stiffness matching, physically valid generation, and a representation that spans diverse topologies. We present TrussOrbit, a structural language model conditioned on nine effective-stiffness components. Its Edge Orbit Truss Representation (EOTR) encodes node geometry and connectivity through Pmmm symmetry orbits. A unified data pipeline consolidates four public sources into 1.49 million canonical structures with common finite-element labels. We introduce Setwise Marginal Credit Optimization (SMCO) to optimize stiffness matching and binary validity across fixed-size candidate groups through sequence-level marginal credit. On the unified benchmark, TrussOrbit achieves the lowest NMAE, NRMSE, and MAPE among the compared models while covering the full dataset; it generates candidates 6.6 times as fast as the fastest external baseline. Representation and policy-objective ablations show complementary gains: EOTR shortens sequences and reduces invalid generation, while SMCO improves stiffness matching over supervised fine-tuning in both single-candidate and best-of-16 evaluation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
inverse design, truss metamaterials, autoregressive generative models, symmetry-aware representations, reinforcement learning, set-level policy optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zU5ROpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work introduces TrussOrbit as a representation of metamaterial that addresses the first issue: the symmetry of metamaterial generation is usually hard to satisfy. Then they introduce an evaluation pipeline SMCO, to evaluate fixed-size groups using stiffness-matching and binary validity rewards. The tokenization method coupled with SFT + SMCO (GRPO-like reward) training enables effective generation of metamaterial.

Critical strengths
  • The generation results are solid, especially regarding the success rate and reconstruction results.
  • The generation latency is state-of-the-art compared to previous work.
Critical weaknesses
  • [Novelty] The introduction claims two existing issues, which are kind of little weak. For the first issue, there are several existing works have emerged to address it by using the property of periodicity and symmetry, such as [1,2]. For the second issue, the one-to-many issue mainly lies in generative model, which can usually be addressed by tuning the generation random seeds or noise, and how this work addresses it is not clear.

  • [Presentations] The writing of main text requires further polish to improve readability.

  • [Presentations] The evaluation requires more details. How the validity check is conducted and how data coverage is computed are not clearly defined.

  • [Results concern] From the qualitative experiments, it seems the model tends to repeat the Target model, resulting in a more similar lower NRMSE. This means lower NRMSE is achieved by repeating the generation; is the model capable of generating new structures or it tend to repeat datasets?

  • [Soundness] The methods use self-defined datasets, metrics, and computation. Some important baselines such as mace+ve [3], and other methods and metrics listed in previous benchmarks [4] can be included to improve the robustness.

Some minor concerns (some suggestions that may not influence rating):

  • Some specialized words require further explanation or definition to improve readability, such as Counterfactual rewards, where the meaning or indication of counterfactual is not clear.

  • Some definitions are not clear. For example, how the v_i is represented; is it a coordinate? A more detailed and formal definition might be better for readers.

  • All notations and abbreviations should be explained when first used, such as FEM, U_im, VGH, MAPE.

  • Paragraph (Lines 243 to 250) is not well-structured. The descriptions of the symbols and the proposed representations are not clear. It might be better to describe the representation design in readable language and explain each symbol accordingly.

[1] uniTruss: Li Zheng, Konstantinos Karapiperis, Siddhant Kumar, Dennis M. Kochmann Nature Communications. 2023

[2] Abu-Mualla, M., Huang, J. Inverse design of curved mechanical metamaterials with geometric AI: a generative diffusion operates in compact latent space of cellular structures. npj Metamaterials..2025.

[3] Grega, Ivan, et al. Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials.ICLR. Vol. 2024. 2024.

[4] Chen et al. MetamatBench: Integrating Heterogeneous Data, Computational Tools, and Visual Interface for Metamaterial Discovery. KDD. 2025

Questions for authors
  • How you explictly address the one-to-many problem and how you evaluate it?
  • Can the generation results be new, or it tends to repeat existing generations with similar stiffness (lower NRMSE)?
  • Is it possible to compare with more methods such as paper "Energy-conserving equivariant GNN for elasticity of lattice architected metamaterials"[3], or more metrics, such as some metrics and methods in the previous metamaterial benchmark[4]?
Reviewer_iSBfOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work presents TrussOrbit, a structural language model conditioned on a variety of stiffness components. they utilize Edge Orbit Truss Representations, Pmmm symmetry orbits, Setwise Marginal Credit Optimization, and sequence-level marginal credit. On the unified benchmark, TrussOrbit achieves the both the best results and the fastest.

Critical strengths
  1. The authors remove duplicated structures, recalculate properties, and separate training and testing structure by topology family. They also decompose and test the edge description and training method separately. Overall, this contributes a high degree of rigor.
  2. Appendix F.2 reports success rates across all targets and counts invalid structures as failures. In my opioin, this gets close to the type of robustness an engineering would want to see in practice.
Critical weaknesses
  1. The main accuracy table in the body of the manuscript omits failed outputs and and error measure includes only structure that hte simulator can evaluate. I also believe that the best-of-16 result uses simulation to choose between candidates, so the reported generation time does not cover the full search time (unless I'm missing something).
  2. One custom beam simulator, ae108, provides the training results, rewards during training, choice among generated designs, and final test scores. Agreement within that system does not show that another simulator or a more detailed structural model would give the same stiffness.
Questions for authors
  1. What is the time needed for simulation during the best-of-16 process?
  2. For a sample of generated designs, how close are the stiffness results from ae108 to an independent structural simulator or a more detailed FEA model?
Reviewer_DzJyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TrussOrbit is a decoder-only Transformer (16 blocks, d = 1024) that generates periodic trusses with Pmmm symmetry. It is conditioned on the nine orthotropic stiffness components, which enter as nine continuous prefix embeddings. The paper has three ingredients: 1) EOTR, a canonical codec 2) SMCO, a reinforcement-learning fine-tuning objective, and 3) A unified dataset.

Critical strengths

TrussOrbit’s strongest contribution is its representation of trusses through symmetry, which produces much shorter sequences while allowing flexible geometry and connections. Combining existing datasets under a common finite-element labeling procedure is also valuable. The representation experiments and the matched direct-reward comparison in Table 15 provide useful evidence that both the representation and the proposed training method improve performance.

Critical weaknesses

The comparisons with existing methods are harder to interpret. The baselines train on different, often much smaller datasets because their representations support fewer structures, and their reported errors are averaged over different subsets of successful outputs. Validity is aso not measured the same way for every method.The retrained TrussGPT reaches 1.18% candidate validity, roughly 817 of the 69,262 test targets. The abstract of the TrussGPT article reports over 99% structural validity. Even with stricter checks, a gap that large means the experiment is measuring the adaptation, not the method. Similarly regarding UniMate, it shows 100.00% validity but a MAPE of exactly 1.0000. That is the value you would get if its FEM responses were (near) zero for every non-zero target component.

I think a comparison on shared training data and structural support is needed to separate improvements from the learning method from those due to access to more data. The representation should also be compared with a baseline that already exploits symmetry.

There is no retrieval baseline and no novelty analysis and no out-of-distribution test. Returning the training structures with the nearest normalized stiffness takes under a millisecond and is always valid, so i think this can be a baseline. Its top-16 version is the natural counterpart to best-of-16.

Use one timing protocol for all methods and report both latency and throughput. FEM cost is excluded even though selection depends on it. Hardware, batch size and inference engine are not reported for any method. Report time per valid result too

No variance or uncertainty is reported anywhere. See every results table (1, 2, 10 to 15).

Questions for authors

See comments in weaknesses. Some may repeat below.

Apply TrussOrbit's own validity predicate to every method. For each baseline, report native validity and error on its original data and protocol, and show they match the published numbers. Break down TrussGPT's failures by cause. Explain UniMate's MAPE of exactly 1.0000. Report top-1 and top-16 retrieval under the same metrics. Report how many generated structures exactly match a training structure (by EOTR sequence) or share its WL4 family. Add an out-of-distribution protocol, for example holding out one source, or using perturbed or extrapolated targets.

How much do SFT, the direct-reward control and SMCO vary across at least three seeds? If the 0.0471 to 0.0432 gap exceeds that spread, my confidence in SMCO rises.

Repeating the supervised baseline, direct-reward control and SMCO under matched conditions would strengthen the paper. A nearest-stiffness retrieval baseline would also help establish what generation adds beyond searching the large labeled dataset.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.