acceptodds
Under review as a conference paper at ICLR 2027

SolidX: Learning Physical Interactions with Structured Bases for Multi-Solid Systems

18 views

Abstract

Simulating contact-induced deformation in multi-solid systems is critical to many scientific and engineering applications, yet remains challenging for neural operators due to the complex and heterogeneous interactions among multiple physical bodies. Transformer-based neural operators have recently emerged as a powerful paradigm for physical modeling, typically compressing high-dimensional physical fields into a shared latent space to efficiently capture long-range physical dependencies. In multi-solid systems, however, such homogeneous latent representations tend to implicitly entangle distinct intra- and inter-solid couplings within the same set of latent tokens. Such entanglement can limit interaction-specific expressiveness, particularly in data-scarce scenarios. Inspired by structured decomposition principles in classical computational mechanics, we propose ***SolidX***, an interaction-structured neural operator that explicitly organizes latent physical bases around distinct interaction patterns in multi-solid systems. By assigning dedicated basis groups to heterogeneous physical interactions, SolidX captures complementary physical behaviors more effectively while retaining linear complexity. Extensive experiments across five multi-solid benchmarks demonstrate that SolidX consistently outperforms strong baselines in prediction accuracy, convergence efficiency, and sample efficiency. Our results indicate that interaction-structured representations provide an effective inductive bias for operator learning, particularly in data-scarce scenarios where high-fidelity data are usually expensive to obtain.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Multi-Solid Systems, Contact-Induced Deformation, Interaction-Structured Representation, Explicit Modelling, Data-Efficient Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_B6GAOpenReview ↗
Rating
3(+1)
Confidence
2(−2)
Informativeness
1
Presentation
2(−1)
Soundness
2
Read review
Paper summary

This paper proposes SolidX, a neural operator for multi-solid deformation that explicitly organizes the representation around physically motivated interaction groups, including loading, continuum, and boundary regions. The method uses group-specific basis functions, masked analysis, hierarchical synthesis, and persistent physical context to preserve interaction structure across layers. The authors evaluate SolidX on five 3D multi-solid deformation benchmarks and report consistent improvements over graph neural networks and neural-operator baselines, including LinearNO and Transolver, across geometry, stress, and PEEQ prediction. They also provide ablations, learning-curve comparisons, effective-rank analysis, and a theoretical formulation of the interaction-structured operator.

Critical strengths

I think the strength of the paper is the consistency of the empirical results. SolidX improves over the compared methods across all reported outputs and across several different physical settings, not like simply showing gains on only one benchmark or metric. The learning-curve results also suggest that the advantage is not limited to the full-data regime. Additionally, the physical inductive bias is also well motivated. Explicitly separating loading, continuum, and boundary interactions is a natural way to encode structure that standard homogeneous neural operators do not represent directly. The formulation is reasonably coherent with this motivation and is supported by useful analyses of the learned bases.

Critical weaknesses

My main concern is that the experiments do not cleanly establish whether the gains come from SolidX's proposed architecture or from the additional physical information it is given. SolidX uses interaction masks, solid identities, loading trajectories, and persistent geometric/interaction context, while the main baselines do not receive an equivalent amount of information. The ablation that removes the proposed grouping does not fully resolve this because it changes the representation and architecture together. A strong baseline such as LinearNO augmented with the same physical masks and conditioning would be important. Without such a control, the central claim about interaction-structured basis representations is difficult to isolate.

I am also not fully convinced that the methodological novelty is sufficiently separated from existing geometry- and slice-based neural operators. At a high level, the main architectural changes are to partition the basis functions by physical groups, apply group-specific masking/projection, and route the resulting representations hierarchically. The current baseline set does not include several relevant operators designed for irregular geometries or structured physical interactions, such as Geo-FNO, DAFNO, GNOT, or Peridynamic Neural Operator. These comparisons would help clarify whether the gains are specific to SolidX's proposed interaction-structured representation or reflect a broader benefit from incorporating geometric or physical structure. The theoretical analysis mostly formalizes the resulting rank and support properties under idealized assumptions, so the empirical comparison is particularly important for establishing the claimed novelty.

Questions for authors
  1. Can you provide a controlled comparison in which LinearNO or another homogeneous operator receives the same interaction masks, solid identities, loading trajectory information, and persistent physical conditioning as SolidX? How much of the reported gain remains?
  2. How are the candidate interaction radius and other parameters defining the physical groups selected? In particular, does their selection use any information from the full test trajectory or ground-truth deformation/contact information?
  3. How robust is SolidX to an incorrect interaction prior? For example, what happens when group assignments are perturbed, noisy, or based on a simpler geometric heuristic?
  4. SolidX only updates deformable points and explicitly masks rigid points. How much of the accuracy and efficiency improvement comes from this additional knowledge, and have the baselines been given the same update mask?
  5. The current interaction construction relies on known rigid loading trajectories. How would the method behave when the interaction structure itself evolves as a consequence of deformation, as in state-dependent contact?
Reviewer_spD3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper provides a way to extend neural operator based physical simulation to contact induced deformation. The basic systems here (illustrated in the appendix) involve two or more soft, solid objects colliding with each other and deforming. For this reason, the system needs to resolve both the internal stresses of the material as well as the contact-based forces. The architecture here uses careful groupings to treat contact, with practical improvement.

Critical strengths

I will preface my review by making it clear that I am far from an expert in this domain; the AC should not wait my review highly, and the authors are welcome to respond or let me know if I'm totally off base.

Contact is indeed a critical thing to get right in physical simulation, and the fact that past neural operators do not carefully treat this effect and separate it (even pointwise) from internal stresses indicates that the overall thrust of this work is worthwhile. The results presented in the paper also seem convincing, in particular the reduced dependence on data relative to past work, suggesting that the proposed architecture has a reasonable inductive bias.

I appreciated Algorithm 1 and the many examples in the appendix, which helped me decipher what's going on (at least broadly).

Critical weaknesses

I'm a generalist with substantial knowledge of algorithms for physical simulation (including contact mechanics), but I have limited knowledge of neural operators. As a reader with this "profile," this paper was an extremely difficult read that made it hard for me to discern the level of contribution.

The text lacks the context needed to understand eq 1,2 in any detail.

The main body of the paper lacks a broad guide to what's going on in the simulation system. For example, the assorted paragraphs in sec 3.3 assume a lot of context that the reader doesn't have (e.g. how is a "contact state" represented?). I would suggest adding figures illustrating the overall pipeline, illustrating typical input/output for the system, and showing what goes wrong with simulations from past work.

The "persistent physical conditioning" is described in a vague paragraph of text without math.

The paper overall comes as a 63 page PDF, so the main body is more of an advertisement for the appendices than a deep exposition of the content. It is not reasonable to ask conference reviewers afforded ~2 weeks to review a group of papers to examine all this content in any detail. So at best, if this paper is published, it will be published as unreviewed/unverified content.

Questions for authors

In the problem formulation in sec 3.1, are the assorted constants (S,N,N_i) determined at training time or at test time? If they can vary at test time, how sensitive is the method to density of sampling relative to the training data?

Isn't the presence of K,V in (1)-(2) indicative of "attention" mechanisms in both architectures? I wasn't sure the role of "Attn" specifically.

Beyond comparing to other ML-based systems for contact simulation, why not compare to recent classical simulation algorithms (e.g. finite elements plus special collision handling) for this task? My understanding is that modern papers (e.g. published at SIGGRAPH or in venues for computational mechanics) include efficient and robust pipelines for this kind of simulation. Unless I missed it, it didn't seem like ML was used to make predictions in the presence of missing information.

Reviewer_mUJ7OpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SolidX restructures slice–deslice neural operators (Transolver / LinearNO) around interaction groups. Candidate contact regions are found from the rigid loading trajectories (KD-tree radius queries) and organized into loading, continuum and boundary groups. Each group gets masked-softmax analysis bases (Eqs. 4–5), and a two-level group/token routing performs synthesis (Eqs. 6–8). Persistent physical conditioning (PPC) and rigid-state preservation (RSP) are also added.

It provides an inductive bias through structural and interaction representations. On five multi-solid datasets with ~100 training samples each, SolidX is best on all 11 outputs (9.7–33.7% gains) and claims ~5× faster convergence and ~80% data savings.

Critical strengths

The design is clean and well motivated. Table 5 reports three-seed results for SolidX and five main baselines. SolidX's standard deviations are small, and most margins exceed the baselines' standard deviations. The ablations and sensitivity studies are broad (Tables 6–9). The experiments are solid and the argument is fairly thorough.

Critical weaknesses

Weaknesses

  1. Input information may be asymmetric. Candidate interfaces are computed from the entire prescribed rigid loading trajectory with a hand-chosen radius. The trajectory is a task input available at inference, not the unknown true contact state, but the derived interface prior is extra structured information. SolidX uses it through group masks and through membership/identity channels in the persistent conditioning. The paper presents this as an inductive bias, but I am concerned that it gives SolidX information that the baselines do not receive. The paper does not say whether the baselines receive the same information. The largest ablation drop (removing grouping → LinearNO) removes both the structure and the information. Please give LinearNO, Transolver and Unisoma the same group-ID / interface-mask channel, and run SolidX with random or shuffled groups. This would separate the contributions of the different components.

  2. The dependence of the efficiency comparison on training budget and convergence is untested. All methods use 100 epochs, batch 1 and lr 3e-4 without per-method tuning. The baseline curves are still falling steeply at epochs 60–100 (Figs. 14, 19–21). Baseline seed variance is large (LinearNO / Tissue 0.168 ± 0.097; Unisoma / Cavity 0.054 ± 0.041). The “5× / 80%” figures are measured against the baselines' 100-epoch error, which would drop with longer training. Still-falling curves alone do not prove that the efficiency advantage comes from under-training; this needs to be tested by training longer. The figures also come from single-run curves and use the most favorable output. Please train to convergence or use official schedules (the official Unisoma Cavity config is 200 epochs, batch 20, 32 slices), with matched learning-rate sweeps. I think this would make the comparison fairer; alternatively, please provide a corresponding FLOPs comparison.

  3. The representation and theory claims are weak. Groups may overlap in practice (p.15), so Corollary 1.1's disjoint-support rank additivity describes only an idealized extreme and does not directly explain the measured effective rank. The masks may nonetheless raise the rank already at initialization; please report rank at initialization and its correlation with error. Prop. 3 counts the dimension of linear hypothesis classes for a fixed input, which does not imply better generalization; the main text should keep the appendix's caveats.

Minor comments

    • The reproduced Unisoma is weaker than its published numbers (e.g., Bilateral PEEQ 0.4097 / 0.3707 vs 0.2817 / 0.2265), which is expected since the original uses 1,000 training samples (Appendix D) versus ~100 here. With SolidX's margins (9.7–33.7%), three seeds and the hyperparameters disclosed in Tables 3–4, this does not affect my assessment. It would still help to run the reproduced baselines several times, and the abstract should state that the results concern the ~100-sample setting. This can be combined with the fairness experiments above.
  • Most Table 1 baselines are single runs; add test-sample bootstrap intervals. This would make the results more credible.
  • SolidX uses the most memory on Bilateral Stamping (Table 10).
  • Clarify the dataset provenance and preprocessing (all five are the datasets of Unisoma's Table 1), and the release plan.
Questions for authors

(1) Do the baselines receive the group / interface information, and what happens when they do? What do random groups give? (2) Does the ranking hold after training to convergence with tuned learning rates? (3) What is the effective rank at initialization?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.