acceptodds
Under review as a conference paper at ICLR 2027

FROST: LEARNING FREE-BOUNDARY EQUILIBRIA FOR SIMULATION AND INVERSE DESIGN

Abstract

Free-boundary problems are partial differential equations (PDEs) whose domain evolves and is not known in advance; they govern solidification, growth and transport across the physical sciences. Classical sharp-interface solvers remesh as the boundary moves, lose robustness when domains merge or split, and are recomputed from scratch for every new instance, while existing neural operators act on a fixed, known domain or track the boundary by a single smooth deformation and so cannot represent changes in topology. We introduce FROST, a free-boundary equilibrium operator that represents the solution as a joint fixed point of the physical field and a level set on one fixed background grid, in which the physical free-boundary condition is the fixed-point residual. Because the geometry is a level set rather than a deformation, a single operator represents merging, splitting and nucleation without remeshing. In its equilibrium realisation the fixed point is solved and trained by implicit differentiation in the manner of a deep equilibrium model, so training memory is constant in the number of solver steps and convergence is monitored; a conditioned realisation of the same operator predicts time-dependent and parametric families in a single pass. The trained operator is then differentiated with respect to a design variable, so forward simulation and inverse design reduce to one problem and the physics residual serves as a trust signal for the design. Across steady and time-dependent free-boundary benchmarks FROST recovers boundaries and the changes in topology that a single deformation cannot represent, learns from only a handful of simulations, and steers a merging front to a prescribed target.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
free-boundary problems, operator learning, deep equilibrium models, level-set methods, inverse design

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jH5eOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper proposes FROST, a neural operator framework for solving free-boundary PDEs and performing inverse design. The method jointly represents the physical field and the evolving domain through a level-set function. FROST introduces two implementations: a deep equilibrium (DEQ) model and a time-conditioned FNO. Experiments on obstacle, tumour-growth, and Stefan problems show the performance of boundary prediction, topology-changing dynamics, and inverse design capabilities.

Critical strengths
  1. Important and relatively underexplored problem. The paper addresses free-boundary PDEs, an important yet relatively underexplored problem in neural operator learning. Unlike conventional PDE problems defined on fixed domains, free-boundary problems require the simultaneous prediction of physical fields and evolving boundaries, potentially involving changes in topology. Developing efficient learning-based approaches for such problems is both scientifically meaningful and practically relevant.

  2. Interesting fixed-point formulation. The paper introduces a fixed-point formulation that jointly models the physical field and the free boundary. This is a conceptually interesting approach, as free-boundary problems naturally involve coupled dependencies between the physical solution and the domain geometry. The use of a deep equilibrium model (DEQ) also enables implicit differentiation without backpropagating through all solver iterations, providing a potentially efficient framework for both forward simulation and inverse design.

  3. Theoretical analysis. In addition to the proposed framework and empirical evaluations, the paper provides a theoretical analysis of the equilibrium formulation. In particular, it establishes the existence and uniqueness of the fixed point, convergence of the iterative solver, and well-defined implicit gradients under a local contraction assumption. These results provide a useful mathematical foundation for the proposed equilibrium-based approach.

Critical weaknesses
  1. Insufficient experimental evaluation and unclear presentation of results. The experimental evaluation is not sufficiently comprehensive. On the one hand, the paper includes only a limited number of baselines and lacks direct quantitative comparisons with several closely related methods discussed in the Related Work section. On the other hand, the experimental results are not presented clearly or systematically. Many results are reported only in the main text or figures, without comprehensive numerical tables summarizing the performance of different methods across evaluation metrics. This makes it difficult to assess the relative advantages of the proposed method and compare the performance of different approaches. I suggest including stronger baselines and providing more detailed quantitative comparison tables.

  2. Unclear motivation for choosing DEQ versus conditioned FNO. The paper adopts different model architectures across experiments: DEQ is primarily used for the steady obstacle problem, whereas conditioned FNO is used for time-dependent problems such as tumour growth and Stefan solidification. However, the motivation behind these choices is not sufficiently explained. What are the respective advantages and limitations of these two architectures, and what types of problems are they best suited for? In particular, it remains unclear whether DEQ offers advantages over conditioned FNO in terms of accuracy, efficiency, or physical consistency. A more detailed discussion of the architectural choices, ideally supported by direct comparisons on the same benchmark, would help clarify the role and necessity of the proposed equilibrium formulation.

  3. Insufficient ablation studies. The paper introduces several components, including the joint field–level-set representation, the fixed-point formulation, and the physics-consistency residual. However, the individual contributions of these components are not sufficiently investigated. For example, it would be useful to compare the proposed level-set representation with simpler alternatives, such as directly predicting the domain mask, and to evaluate the benefits of the equilibrium formulation over a standard feed-forward neural operator under comparable settings. More comprehensive ablation studies would help establish which components are essential and how much each contributes to the overall performance.

  4. Limited evaluation of different types of topology changes. Although the paper emphasizes the ability to handle general topology changes, including merging, splitting, and nucleation, the experiments primarily focus on merging events. The ability to represent arbitrary topology changes using a level-set function does not necessarily guarantee that the learned operator can accurately predict such changes. Additional experiments involving splitting, nucleation, or more complex topology transitions would provide stronger evidence for the generality and robustness of the proposed approach.

Questions for authors

See above.

Reviewer_d2kwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a neural operator framework for solving free-boundary partial differential equations and performing inverse design in settings where the computational domain can change topology. The central idea is to jointly represent the physical field and the free-boundary geometry through a level-set function on a fixed background grid, allowing interfaces to merge or split without remeshing.

The framework includes two realizations: (1) a DEQ operator that solves a learned fixed-point equation and uses implicit differentiation, and (2) a conditioned feedforward spectral operator for time-dependent and parametric problems. The authors also propose physics-consistency residuals as inference-time trust indicators and use operator differentiation for inverse design.

Validation covers three two-dimensional benchmarks: a steady obstacle problem, time-dependent tumour growth, and Stefan solidification. The reported results include high interface accuracy and successful surrogate-based inverse design of merging solidification fronts. The paper additionally presents a local-contraction convergence theorem and an inverse-design experiment validated using an independent solver for the obstacle problem.

Critical strengths
  1. propose a formulation for learning free boundaries with changing topology.
  2. develop an integration of forward prediction, implicit differentiation, and inverse design.
Critical weaknesses
  1. The claimed equivalence between the learned fixed-point residual and the free-boundary condition is insufficiently established.
  2. The experimental evaluation does not adequately isolate the benefit of the equilibrium formulation or independently validate the main topology-changing design results.
Questions for authors
  1. Clarification of the physical interpretation of the fixed point.
  2. Validation of the equilibrium formulation and comparison with topology-capable baselines.
Reviewer_dazVOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FROST jointly predicts a physical field and a level set on a fixed background grid to represent an unknown boundary and changes in its topology. The steady formulation uses a deep equilibrium model and implicit differentiation, while the time-dependent formulation uses a spectral network conditioned on time. The trained models are further used for inverse design. The paper reports boundary prediction, topology recovery, and design results on obstacle, tumour coalescence, and Stefan solidification problems, and proposes a physics-consistency residual as an auxiliary indicator of prediction reliability.

Critical strengths
  1. Jointly representing unknown boundaries and physical fields through a level set is well motivated and avoids the structural limitation that a single invertible deformation cannot change connectivity. The paper demonstrates merging events across several tasks. On the obstacle problem, the improvement from using an explicit level set over extracting the boundary by thresholding the predicted field also supports the practical value of this representation.

  2. On the steady problem, the DEQ implementation achieves errors of the same order of magnitude as the direct model with fewer parameters, demonstrating a meaningful trade-off between accuracy and model size. The obstacle inverse designs are also recomputed using the reference solver, providing an independent basis for assessing the reliability of surrogate-based design.

Critical weaknesses
  1. The central claim that the free-boundary condition is the fixed-point residual is not sufficiently supported. This equivalence holds for the projected obstacle iteration presented in the paper, but it is unclear how the actual learned latent cell and its decoder preserve this property. The appendix explicitly distinguishes the latent solve residual from the physical residual evaluated on the decoded state. Supervised fitting and numerical convergence under a local contraction assumption do not, by themselves, establish satisfaction of the physical equations, boundary conditions, or field–geometry consistency. The paper should clarify which constraints are enforced by the architecture and which are only approximated through data, and reconcile the residual definitions in the design algorithm and the main text. The authors acknowledge that the Stefan phase–temperature check is only a necessary condition. A low value can serve as an auxiliary diagnostic, but cannot independently validate heat balance or the interface motion law.

  2. The evidence from the time-dependent benchmarks and inverse designs is insufficient to support the physical prediction claims. The supplied tumour data-generation procedure appears to construct geometric merging trajectories first and then solve the nutrient field frame by frame. This is not equivalent to evolving the boundary according to the current nutrient flux, as described in the main text, and its correspondence to the reported experiments needs confirmation. The Stefan time window also appears to depend on the merging process of each instance; the normalization of the time input and its implications for interpreting control effects require clarification. The authors disclose that the Stefan designs are evaluated only through the surrogate, but independent solver validation on the obstacle problem does not directly validate these time-dependent designs. A small set of optimized controls should be recomputed independently. In addition, the existing comparisons support the value of the topology representation and parameter efficiency, but citing training sample counts across different problems does not establish the claimed hundredfold improvements in data efficiency.

Questions for authors
  1. Which physical constraints are enforced directly by the equilibrium cell or decoder, and which are only approximated through supervised learning? Which residual is used in the design acceptance criterion? If the residuals are not equivalent, please qualify the claim that the free-boundary condition is the fixed-point residual.

  2. Please confirm whether the data used in the tumour experiments satisfy the stated flux-driven interface motion law, and explain the relationship between the Stefan time input and physical time. If prescribed geometric trajectories or instance-dependent time normalization are used, how do these choices affect the conclusions about dynamics prediction and control effects?

  3. Could a small set of optimized Stefan designs be recomputed using the reference solver, with target agreement and relevant physical errors reported under the same definition of time, to assess the actual outcome of surrogate-based optimization?

  4. Is the claimed hundredfold improvement in data efficiency supported by comparisons on the same task under the same accuracy criterion? If not, please narrow this quantitative claim based on comparisons across tasks.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.