acceptodds
Under review as a conference paper at ICLR 2027

Homomorphic Generator Learning: Compositional PDE Neural Operators for Multi-Physics Surrogates

Abstract

PDE surrogates may be asked to predict systems whose physical mechanisms were observed separately, but never together, during training. Although the corresponding equation terms add, their solutions generally do not; a learned solution map therefore has no explicit rule for combining those observations. We propose (HGL), which learns a reusable generator for each mechanism and constructs an unseen system by adding its active generators. Each generator combines a role-constrained principal operator with a gated correction. Known symbols provide operator priors; otherwise, the symbols are learned from data. Mechanism coefficients are supplied by the equation or estimated from a short observed trajectory. A task-specific evaluator converts the combined generator into a rollout or steady-state solution. On unseen advective Fisher–KPP dynamics, HGL improves long-horizon prediction with either supplied or estimated coefficients relative to the evaluated solution-operator baselines. It also performs well as training coverage expands from individual mechanisms to partial combinations and when generators learned from different equation families are reused in a steady-state problem. These results support generator-level composition as an inductive bias for multi-physics generalization.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural Operators, Operator Semigroups, Operator Theory, Scientific Machine Learning, Multi-Physics Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_P7Q1OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper focuses on Neural Operators and combining different PDE differential operators. Given PDEs that are defined through a single transport, diffusion, or generic differential term, the objective is to learn a model that, trained on these individual instances, can combine the mechanisms defining their dynamics, leading to a model that is capable of solving the PDE with all terms active. The key idea is to consider a generator operator for each term and combine them via an appropriate algebra. The results show that the model is indeed capable of combining differntial terms to unseen combinations.

Critical strengths
  • The problem is relevant and interesting, with few approaches that have attempted to solve it. The problem definition is also sound, with good motivation, and the results achieved are great.
  • The proposed solution seems sound: trajectories generated from individual mechanisms cannot be trivially combined, and it is thus necessary to combine the mechanisms with a meta approach.
Critical weaknesses

The paper has several issues, from the presentation of the methodology to the experiments:

  • The related works section lacks too many details and feels more like a list than a coherent comparison of models. In particular, works such as [1] and [2] and [3], which constitute some of the main competitors, are only sketched.
  • Furthermore, these related works try to solve the same problem as the paper, that is, combining seen physical mechanisms on unseen combinations or, more generally, combining physical operators. However, there are no experimental results concerning these models. The models HGN is compared to, such as U-Net and FNO, are all designed to solve a single specific problem and do not naturally extend to unseen physics. It is also not clear how these models are trained and then tested on different physics at all. In the current setting, the experiments are not sound.
  • The whole paper, and especially Sections 3 and 4 (the main contributions), are not coherently written and are very hard to follow. Some objects are never defined or are used before they are defined. The presentation is very schematic and lacks many details, especially examples and interpretations that are key to understanding new concepts. For example, what is a "state-dependent tendency map" (line 143)? To the best of my knowledge, "tendency" does not refer to a common concept in this field. Also, it is not clear what "medium" (line 152) refers to. The major problem is that these are only examples of a larger problem with the whole presentation.
  • While the model seems to be theoretically sound, no theoretical claims support how the composition of operators works, apart from a brief and not really understandable discussion in Section 4.1.
  • Some references are wrong. For example: Hmida et al., Compositional Neural Operators for Multi-Dimensional Fluid Dynamics, is not from the main track of ICLR 2026 but from a workshop

[1] Louis Serrano et al. "Test-time generalization for physics through neural operator splitting". ICML 2026

[2] Vignesh Gopakumar et al. "Learning physical operators using neural operators." AISTATS 2026

[3] Harde et al., "Poseidon: Efficient foundation models for PDEs", NeurIPS 2024

Questions for authors
  • Can the authors provide a more in-depth comparison to [1,2,3], both from a modeling and experimental point of view?
  • The authors should provide more details on the presentation of the model and the composition of operators, define objects in order, and add clear physical interpretations.
  • Are there any theoretical claims further supporting the model and how it works for PDEs?
3 comments

W1: The section reads as a list, and the main competitors [1, 2, 3] are only sketched.

Thank you for this suggestion. Following it, Section 2 of the revised paper now includes the detailed comparison with [1], [2] and [3]. HGL shares with [1] and [2] the additive decomposition of the right-hand side into coefficient-scaled terms. In addition, each learned term is constrained to the differential order of its mechanism, which separates mechanisms observed only in mixtures. The terms are added before evaluation, so the same composition serves evolution and steady solves, and known operators enter directly alongside learned terms. Operator splitting [1] composes pretrained operators selected by a search over a trajectory of the test system. The approach of [2] learns right-hand-side terms next to known linear terms and integrates their sum. Poseidon [3] is a pretrained solution operator that learns the flow map.

W2, Q1: There are no experimental results against [1, 2, 3], and it is unclear how the baselines are trained and tested on unseen physics.

Thank you for this suggestion. As you asked, the revised paper now includes this experimental comparison. On training systems, HGL, the physical-operator model, LegONet, and CompNO fit to relative errors below 1%. On unseen combinations, HGL is 15 to 28 times more accurate. Every baseline is trained on the same training systems, receives the same state history and mechanism coefficients, and is tested on the unseen combination (Section 5.1 of the revised paper).

Five-mechanism, single-mechanism training, step 60Training systemsUnseenError ratio (baseline / HGL)
HGL (order-constrained)0.00080.0034—
Physical-operator model ([2])0.00130.09528×
LegONet0.00130.05015×
Operator splitting ([1])0.00160.07422×
Generic MLP terms ([2])HGLError ratio (baseline / HGL)
Mixture, unseen mixing ratio (step 60)0.160.009417×

These models share HGL's known operators and the same network per mechanism term, without the order constraint. They differ in training objective, in whether the terms are summed or applied by operator splitting, and in whether the mechanism coefficients are given or estimated from four context frames (Appendix D, Table 8 of the revised paper). In Table 2 of the revised paper the baselines also receive the three instantaneous mechanism tendencies (the contribution of each mechanism to ∂_t u) as input. HGL can predict the unseen system from an initial state alone. Operator splitting [1] requires a trajectory of the test system. Poseidon [3] is a pretrained solution operator, the class of the seven solution-operator baselines. These results are in Section 5.2 (Tables 4 and 5) and Appendix D of the revised paper.

W3, Q2: Sections 3 and 4 use objects that are not defined, for example "tendency" and "medium".

Thank you for these comments. Following your comments, the revised paper defines these objects in Sections 3 and 4. Section 3 defines the tendency of a mechanism as its contribution to ∂_t u, and the medium c as the spatial coefficients, material fields or geometry. Section 4.2 gives the explicit form of each learned term, with diffusion, transport, reaction and forcing as examples.

W4, Q3: No theoretical claim supports how the composition of operators works.

Thank you for this question. In response, Section 4.2 of the revised paper states that an unseen combination can be predicted if and only if its coefficient vector is a linear combination of the training coefficient vectors. When two mechanisms always appear with a fixed mixing ratio, the data determine only their sum, and the order constraint determines how the sum is divided between them. Table 5 confirms each case. The first result follows from the estimability theory of linear models (Scheffé, 1959), and the second from the identifiability theory of additive models (Hastie and Tibshirani, 1990). Both references are cited in Section 4.2 of the revised paper.

W5: The CompNO reference is wrong.

Thank you for pointing this out. The CompNO entry now cites the ICLR 2026 workshop in the revised paper.

References: [1] Operator splitting (Serrano et al., 2026). [2] Physical-operator model (Gopakumar et al., 2026). [3] Poseidon (Herde et al., 2024). Scheffé (1959), The Analysis of Variance. Hastie and Tibshirani (1990), Generalized Additive Models.

I thank the authors for the detailed response and the changes to the manuscript. However, there are substantial changes to the paper and methodology that concern me. These changes have also not been highlighted enough (even the coloring in the revised manuscript is not consistent with the changes) and raise additional comparison issues.

  • Previously, the individual operators that the method composes in the final operator were somewhat general, with some physics-induced restrictions. In the revised version, the parameterization of the individual operators is much more restricted, to the point where they are tailored to the specific operator they are approximating, which is also very close to Gopakumar et al., 2026. (see next point). Sections 4.2 and 4.3, on this topic, have lots of additions that have not been properly highlighted. Most importantly, the five-mechanism experiment results changed dramatically (without highlighting them): the error at 60 steps has reduced from 0.385 to around 100 times lower (0.0034), and the number of parameters changed from around 1.6 million to 0.02 million. These potential changes in results and methodology worry me, and it is unclear if things have changed in the two versions.
  • Most importantly, and directly related to the previous point, the paper is now much closer to the already mentioned baseline Gopakumar et al., 2026. In their work, the PDE is decomposed into individual operators, which are implemented analytically or learned depending on their type (linear or nonlinear). This still seems like the current description of HGL, which poses some additional constraints, but now the novelty needs further clarification.
  • The new experimental comparison with LegONet and CompNO needs clarification, especially given the above points. The experimental setup in Appendix D states that these baselines have the same architecture and evaluator as HGL. For example, CompNO originally used Neural Operators to learn nonlinear operators, which are combined by a learnable final update. In this manuscript, the mechanisms seem to be approximated with MLPs. The authors should detail whether the implementation they used is the same as the original work or if their adaptation changed it substantially and, in that case, reevaluate by keeping the methods as close to the original ones, or explain if this is not possible.
  • In general, many of the numerical results in tables appear to have changed, sometimes substantially, and it is not clear where these changes originate.

Overall, I am quite concerned by the changes in the paper and how these have been highlighted, and I expect more clarification from the authors, especially regarding the similarities with Gopakumar et al., 2026, possible changes in the methodology and in the results.

I remain available for further clarification from the authors, in case there have been misunderstandings or to highlight the changes in the paper.

Thank you for the review and comparison. We apologize for the incomplete highlighting in the previous version, which was meant to emphasize the key points. The updated PDF highlights every changed table entry and every sentence with new content (Tables 5, 8, and 15 in full).

The method is the same as in the original paper. Following your comments, the revision makes the order constraint explicit and applies it to every learned term, and the five-mechanism error falls from 0.385 to 0.0034.

Comment 1: The operators are more constrained, tailored to the specific operator each approximates.

Thank you for this comment. The "physics-induced restrictions" you identified in the original paper and the order constraint in the revision are the same restriction. It sets the differential order of each learned term, and a generic pointwise MLP learns the function within that class from data. In Section 4.1 of the original paper, the principal symbol "sets the leading spatial order," and when the operator is unknown, "a role-constrained symbol network" learns it. Section 4.2 grouped the principal terms into "three role-specific families" (diffusion, transport, reaction/forcing). The revision names this the order constraint, writes each term's explicit form (first-order −a(u)·∇u, zeroth-order u·g(u), forcing a free grid field), and states when the terms are identifiable. The logistic rate 1−u is learned from data, and the same classes serve the Navier–Stokes and three-dimensional systems (Table 15).

Comment 2: The paper is closer to Gopakumar et al., and the novelty needs further clarification.

Thank you for raising this. HGL shares with Gopakumar et al. the additive right-hand-side decomposition and the distinction between analytic and learned terms. In addition, HGL constrains each learned term to its mechanism's differential-order class, states when the terms are identifiable (Section 4.2), and uses the same dictionary for evolution and steady-state solves (Section 5.3). The original paper already had this constraint, and the revision names it and makes it prominent. When two mechanisms are observed only together at a fixed mixing ratio (Table 5), generic MLP terms (the Gopakumar et al. design) give 0.16 on the unseen ratio and order-constrained terms give 0.0094, 17 times lower. On the unseen five-mechanism system (same data and training budget, Table 8), generic MLP terms give 0.095 and HGL gives 0.0034, 28 times lower.

Comment 3: The five-mechanism error changed from 0.385 to 0.0034, parameters from 1.60M to 0.02M, and many other table entries changed.

Thank you for the detailed check. Following your comments and those of Reviewers auT4, 8AGa, and Gjea, the revised five-mechanism experiment gives each learned mechanism its own order-constrained term following Sections 4.1 and 4.2 of the original paper, and the error falls from 0.385 to 0.0034, about 100 times lower. In the original code, the three learned mechanisms shared one network (1.60M parameters), an implementation issue in this experiment. The best baseline remains at 0.662. Eq. (1), data, training budget, evaluation, and all baseline rows are unchanged.

TablesChangeSource
4, 10–13, HGLStep-60 error 0.385/0.318/0.179/0.193 → 0.0034/0.0050/0.0068/0.0069; params 1.60M → 0.02MOne order-constrained term per learned mechanism (above).
1–2, HGLHGL params 2.03M→1.53M; step-80 error 0.00036→0.00031. DPOT params in Table 1 2.22M→1.54M (errors unchanged)HGL sized below the smallest baseline (1.54M), three seeds. DPOT count matches Table 2.
3Single seed → three-seed medianRequested by Reviewers Gjea and 8AGa.
913 of 18 rollout errors shift at 3rd digitRecomputed with 1.53M model.
5, 8, 15New experimentsReviewer requests; Table 8 includes CompNO with FNO modules per your suggestion.
All baseline rows; 6, 7, 14Unchanged—

Comment 4: The LegONet and CompNO baselines use HGL's architecture and evaluator and should stay as close to the original methods as possible.

Thank you for this suggestion. Following it, Table 8 includes CompNO with per-mechanism FNO modules and a learned aggregator, and on the unseen five-mechanism system at step 60 the error is 0.079 (HGL: 0.0034, 23 times lower). With the aggregator trained on target-system trajectories as in the original CompNO, the error is 0.044 (13 times lower). LegONet trains by operator matching with labels from a trusted discretization, unavailable when mechanisms are unknown, so its row keeps the training objective and Strang splitting with generic MLP terms. Appendix D describes each row's relation to the original method.

Thank you again for the thorough check.

Reviewer_8AGaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper introduces a modular alternative to conventional PDE surrogates by learning a separate, lightweight generator for each physical mechanism and combining these generators to model new systems. Its central claim is that mechanisms learned independently can be composed to generalise to previously unseen combinations. The approach is evaluated on the Fisher–KPP equation, a five-mechanism system, and a steady-state convection–diffusion–reaction problem.

Critical strengths

The Fisher-KPP experiment is the strongest part of the paper. Each mechanism appears alone during training, and the model has to combine all three at test time. The resulting error is orders of magnitude lower than FNO and the other baselines in Table 1, and since the parameter counts are comparable, the gap isn't a capacity effect. This experiment directly tests the paper's central claim: that mechanisms seen only in isolation can be reused in unseen combinations. A second strength is the separation between adding generators and evaluating the sum. The paper explains why this separation is necessary and "Neither response is generally a sum of separate responses." So the same dictionary can be reused across tasks, and only the evaluator needs to change.

Critical weaknesses

The experiments cover a narrow range of PDEs. Everything is two-dimensional and built from simple terms: transport, diffusion, reaction, forcing, and the nonlinear reaction in Fisher-KPP. The steady problem is linear, and nothing here explore Navier-Stokes, high Reynolds numbers, three dimensions, or coupled fields, even though the title and abstract read as general.

It is not clear how much of the method depends on analytic symbols. On Fisher-KPP, the additive model with analytic symbols beats every baseline by a margin in Table 1. Swap in learned symbols and the error rises by about three orders of magnitude in table 3. Replace given coefficients with fitted ones and it rises by about seventy times .

On The Well, which uses learned symbols throughout, HGL is best in six of nine settings in Table 6. The paper does not give a clear account of when learned symbols hold up, and the abstract's "analytic priors where available" does not tell the reader that they might not.

Table 3 supports two of the paper's main claims: additive composition beats nonlinear composition, and learned symbols perform poorly. But Table 3 uses one seed and two hundred training steps, while the main results use three seeds. Those two claims should be re-checked.

Questions for authors

(1) The title says the method covers PDE mechanisms, but the experiments are all two-dimensional and low-order. How would the method behave on Navier-Stokes or at higher Reynolds numbers, and how much of the two-dimensional result carries to three dimensions or higher dimensions?

(2) Learned symbols underperform on Fisher-KPP (Table 3) but are used throughout The Well, where HGL wins in several but not all settings (Table 6). What distinguishes the cases where learned symbols work from those where they fail?

(3) Table 3 uses one seed and a short training run, while the main results use three seeds (Section 5.1). Can the ablation be reproduced with three seeds and a comparable budget, so that the ordering between additive and nonlinear composition is shown to be stable?

(4) The static evaluator builds its preconditioner from the problem structure, but page 4 states that the dictionary is shared across configurations. How far does that sharing extend when the evaluator must be rebuilt for each problem?

1 comment

W1, Q1: The experiments are two-dimensional and built from simple terms, while the title and abstract read as general. How does the method behave on Navier–Stokes or in three dimensions?

Thank you for this question. Following your question, the revised paper adds a three-dimensional system and a Navier–Stokes case, and HGL composes learned mechanism terms in both.

New experiments in the revised paper (error at step 60 on the unseen combination)HGLKnown operators only
Three-dimensional five-mechanism system, trained on single-mechanism systems0.000480.25
Two-dimensional Navier–Stokes at Reynolds number 1000, with learned damping and forcing0.0151.25

In the Navier–Stokes case, advection and viscosity are known, and a nonlinear damping and a forcing are each learned from systems that contain only one of them. The error of 0.015 is close to that of the same time integrator with the true terms (0.012). These results are in Section 5.2 and Appendix E.3 of the revised paper. The remaining composition experiments use two-dimensional scalar systems, so that each mechanism and its coefficient can be controlled individually, and The Well adds three multi-field systems.

W2, Q2: It is unclear how much the method depends on analytic symbols, and when learned symbols work.

Thank you for this question. As the physics provided to HGL decreases, the advantage over the strongest baseline persists: 700 times with analytic symbols and coefficients, 8 times when coefficients are estimated from data, 28 times with three of five mechanisms learned, and 17 times from mixture data alone. All baselines receive the same information. Learned terms compose when the training data or the order constraint separates them (Section 4.2 of the revised paper).

Known physics supplied to HGLTraining systemsUnseenStrongest baseline (unseen system)Error ratio (baseline / HGL)
Analytic symbols (Fourier multipliers) and coefficients (Fisher–KPP, step 80)—0.000310.22700×
Analytic symbols, coefficients estimated from data (Fisher–KPP, step 80)—0.0270.228×
Three of five mechanisms learned, single-mechanism training (step 60)0.00080.00340.09528×
The same mechanisms learned only from mixtures (step 60)0.000710.00940.1617×

The strongest baseline is the best solution operator (first two rows) and the physical-operator model (last two). The key structural choice is learning the right-hand side term by term and integrating. Known mechanisms enter through their analytic form. The order constraint on each learned term drives the composition. On combined-system states, the order-constrained terms stay within 1–2% of the true tendency (0.009 and 0.011 for advection and reaction), while the generic MLP terms are 25–30% off (0.294 and 0.245). LegONet, CompNO, and operator splitting fit the training trajectories to comparable accuracy (0.0013 to 0.0059) and give 0.050 to 0.090 on the unseen system (Appendix D, Table 8).

Removing the order constraint in Table 3 raises the error to 0.40, against 0.00031 for HGL. On three multi-field systems from The Well with no analytic operators, evaluated under three protocols each (nine settings), HGL attains the lowest error in six, with reductions up to 54%, and within 12% of the best baseline in the other three (Section 5.4 of the revised paper).

W3, Q3: Table 3 uses one seed and 200 training steps. Is the ordering stable with three seeds?

Thank you for this suggestion. Following it, Table 3 of the revised paper uses three seeds at the 200-step budget of Table 1, and the ordering remains stable.

Table 3 of the revised paper, unseen system, step 80error
Additive (HGL)0.00031
Residual nonlinear composition (sum + MLP correction)0.0096
Nonlinear composition (MLP combining mechanism terms)0.43

Q4: The static evaluator builds its preconditioner from the problem, while the dictionary is said to be shared. How far does the sharing extend?

Thank you for this question. The mechanism operators are shared, and the evaluator is one fixed construction that is never retrained. In the steady problem, the diffusion, transport and reaction operators are trained on diffusion, convection–diffusion and screened-Poisson equations and are added for the unseen equation with all three terms. The preconditioner is computed from the coefficients of each problem by the same formula, and the iteration is identical for every equation family (Appendix C of the revised paper). The time-dependent and steady models are trained separately. What they share is additive composition.

References: LegONet (Zhang et al., 2026), physical-operator model (Gopakumar et al., 2026), CompNO (Hmida et al., 2026), operator splitting (Serrano et al., 2026).

Reviewer_GjeaOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes HGL, which learns one generator per PDE mechanism (diffusion, transport, reaction, and so on). It composes an unseen multi-mechanism system by summing coefficient-scaled generators, then passes the sum to a task-specific evaluator (time integrator or preconditioned iterative solve). Experiments cover advective Fisher–KPP after singleton-only training, a five-mechanism system under increasing partial-combination coverage, a steady CDR problem with the transport+reaction pairing held out.

Critical strengths
  • The held-out-combination protocol is a clean and useful way to separate reuse of known terms from discovery of new ones.
  • The paper correctly separates additivity at the level of the tendency from non-additivity at the level of the solution. The writing is clear about where the composition law lives.
  • The reproducibility artifact and protocol details are good.
Critical weaknesses

Weaknesses

  1. The core idea is close to trivial. Eq. 1 is linear in θ, so the "semigroup homomorphism" (Eq. 3, Eq. 10) holds by construction. It is the method of lines with a dictionary of right-hand-side terms. The algebraic framing adds vocabulary, not content.

  2. The closest related work is not evaluated. Gopakumar et al. (2026), which learns nonlinear terms, fixes the linear terms, and integrates as a neural ODE, is nearly the same construction. Serrano et al. (2026, operator splitting) and CompNO also target exactly this compositional setting. Instead, all comparisons are against monolithic solution operators, which the protocol disadvantages by design.

  3. The ablations undercut the stated contributions.

    • Table 3 uses one seed and 200 steps.
    • Unconstrained symbols beat role-structured ones, so contribution (2)'s structure claim is unsupported, as the authors concede.
    • More importantly, "learned symbols" with additive composition fails (0.4–0.5 error). The operative inductive bias is therefore the analytic prior, not additivity. This also conflicts with the five-mechanism experiment, where 3 of 5 generators are learned from data and HGL still wins. The discrepancy needs an explanation.
Questions for authors
  • What does HGL learn in Table 1 beyond the analytic terms? Please give an ablation with ρ_m ≡ 0.
  • What is the performance against Gopakumar et al. and PDE-FIND + solver under the same protocol?
2 comments

W1: Eq. (1) is linear in θ, so the homomorphism holds by construction. The method is a method of lines with a dictionary of right-hand-side terms.

Thank you for this comment. HGL shares the additive right-hand-side structure you describe with the physical-operator model, LegONet, and CompNO. In addition, each learned term carries the order constraint, which is what makes the terms reusable: with exact summation alone, the error on the unseen mixing ratio is 17 times higher (Table 5 of the revised paper).

Following your comment, the revised paper adds an experiment that tests this directly (Table 5 of the revised paper). With exact summation alone, the error on the training mixing ratio is 0.0014 and on the unseen one 0.16. With the order constraint, HGL gives 0.0094, 17 times lower. Reaction and forcing are observed only in mixtures with a fixed mixing ratio: the sum of the two mechanisms is learned correctly, but its division between them is not determined by the data.

Following Reviewer P7Q1's request for a theoretical statement, Section 4.2 of the revised paper states when each learned term coincides with the true mechanism, and Table 5 of the revised paper confirms each case.

Two mechanisms observed only in mixtures (error at step 60)Training mixing ratioUnseen mixing ratioIdentifiable quantity
No order constraint, fixed mixing ratio in training0.00140.16only the sum
Order-constrained, mechanisms of different differential orders0.000710.0094each term, up to a constant
Order-constrained, mechanisms of the same differential order0.00110.19only the sum
No order constraint, mixing ratio varied in training0.00140.0017each term

The order constraint recovers each mechanism where the data alone give only their sum, which supports the structure claim of contribution (2).

W2: The closest work (Gopakumar et al., operator splitting, CompNO) is not evaluated. All comparisons are against monolithic solution operators.

Thank you for pointing to this work. Following your comment, the revised paper evaluates all three. On training systems, HGL, the physical-operator model, LegONet, and CompNO fit the single-mechanism training trajectories to relative errors below 1%. On the unseen five-mechanism system, HGL gives 0.0034, against 0.050–0.095 for the alternatives (15 to 28 times lower).

Five-mechanism, single-mechanism training, step 60Training systemsUnseenError ratio (baseline / HGL)
HGL (order-constrained)0.00080.0034—
Physical-operator model0.00130.09528×
LegONet0.00130.05015×
CompNO0.00590.09026×
Operator splitting0.00160.07422×

The gap comes from whether each learned term captures its own mechanism. Summing the mechanism terms and applying them one after another by operator splitting give the same result (0.095 in both cases). Training each term to match the time derivative (LegONet) improves the generic MLP terms from 0.095 to 0.050, but they remain 15 times above the order-constrained terms (0.0034). These models share HGL's known operators and the same network per mechanism term, without the order constraint. They differ in training objective, in whether the terms are summed or applied by operator splitting, and in whether the mechanism coefficients are given or estimated from four context frames (Appendix D, Table 8 of the revised paper).

On the mixture task, HGL gives 0.0094, against 0.16 for generic MLP terms (17 times lower; Table 5 of the revised paper).

W3: Table 3 uses one seed. Unconstrained symbols beat role-structured ones. Learned symbols with additive composition fail, which conflicts with the five-mechanism experiment.

Thank you for these observations. As you and Reviewer 8AGa suggested, Table 3 of the revised paper reports three seeds. Additive composition attains the lowest error on the unseen system, and unconstrained symbols (the analytic operators multiplied by free complex factors) no longer beat the role-structured (order-constrained) ones (0.00035 against 0.00031).

Table 3 of the revised paper, unseen system, step 80error
Additive (HGL)0.00031
Unconstrained symbols (analytic operators × free complex factors)0.00035
Residual nonlinear composition (the sum plus a zero-initialized MLP correction)0.0096
Nonlinear composition (an MLP in place of the sum)0.43

Regarding the structure claim of contribution (2): the order constraint matters when mechanisms need to be separated from mixture data, with HGL giving 0.0094 against 0.16 for generic MLP terms (17 times lower; Table 5 of the revised paper). In this ablation all three mechanism operators are specified analytically, so the order constraint does not change the result.

Regarding whether additivity is the operative inductive bias: the operative bias is exact summation together with the order constraint. Without exact summation (the sum plus an MLP correction), the observed systems are still fitted (MSE 2.8×10⁻⁵) and the unseen error rises to 0.0096, 30 times higher (Table 3 of the revised paper). Without the order constraint, the error rises to 0.40.

Regarding the learned-symbol row of Table 3 and its consistency with the five-mechanism experiment: the two results are consistent, because this row removes the order constraint (replacing the analytic operators by free per-frequency parameters), while the five-mechanism experiment retains it. Without the order constraint, the error is 0.40, against 0.00031 for HGL. With order-constrained terms, the learned mechanisms compose at 0.0034 (0.385 in the original paper), against 0.095 for the physical-operator model (28 times lower). Prompted by your comment on this discrepancy, the revised five-mechanism experiment has a more suitable configuration: transport and diffusion are given as known operators, and three other mechanisms are each learned from single-mechanism trajectories.

Q1: What does HGL learn in Table 1 beyond the analytic terms? An ablation with ρ_m ≡ 0 is requested.

Thank you for suggesting this ablation, which the revised paper now reports in Appendix D. In Table 1, the three mechanisms of Fisher–KPP are all known, so the model uses their exact operators and the learned correction is not needed. This is the same use of known operators as in any numerical solver. Removing the correction (ρ_m ≡ 0) gives the classical operator-splitting solver, with errors of 7.7×10⁻⁶ at step 10 and 3.5×10⁻⁵ at step 80 (the best solution operator gives 0.50 and 0.47). The learned components of HGL are evaluated in Tables 4 and 5 of the revised paper, where three of the five mechanisms are learned from data.

Q2: What is the performance against PDE-FIND with a solver under the same protocol?

Thank you for suggesting this comparison, which the revised paper now includes. On unseen combinations, HGL is 180 to 280 times more accurate than PDE-FIND with a solver, including for reactions outside any polynomial library. PDE-FIND identifies a single observed system. HGL predicts a combination that was not observed.

PDE-FIND + solverHGLError ratio (baseline / HGL)
Mechanisms observed only in mixtures, unseen mixing ratio (step 60)0.210.00075280×
Same, with a reaction outside the library, u/(1+|u|)0.230.0013180×
Same, with u/(0.1+|u|)0.320.0012270×

On a single observed system whose terms lie in its library, PDE-FIND is accurate: 0.0039 at step 80 on Fisher–KPP from four context frames (HGL gives 0.00031). HGL can predict a system of which no trajectory is available, from its coefficients and an initial state alone. In the revised paper, Section 2 discusses this comparison, and Appendix D and Table 5 report the numbers.

References: LegONet (Zhang et al., 2026), physical-operator model (Gopakumar et al., 2026), CompNO (Hmida et al., 2026), operator splitting (Serrano et al., 2026), PDE-FIND (Rudy et al., 2017).

Reviewer_auT4OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Homomorphic Generator Learning (HGL) to address compositional generalization problem for PDE surrogates. HGL keeps a mechanism generator dictionary whose elements are role-constrained principal operators. Therefore, unseen combinations can be constructed by summing the relevant generators, scaled by coefficients which are given or estimated from the short context, and the summed generator is then converted into a rollout or steady-state solution by a task-specific evaluator.

Critical strengths

The problem setup is clear and well motivated. Every mechanism (e.g., transport, diffusion, reaction) is seen during training, and only their full combination is new at test time. The evaluation therefore targets compositional generalization itself, rather than discovery of new mechanisms.

Critical weaknesses
  1. Baselines are not enough to show the proposed method’s contribution. For example, in the main Fisher-KPP experiment (cf. Section 5.2), HGL receives the analytic operators for all three mechanisms and even the true coefficients for vanilla HGL from Tables 1 and 2. Therefore, it doesn’t seem like a fair comparison which can show its contribution. Instead, the comparison should be done with the physics-based hybrid operators (e.g., [1], [2]), where the same amount of information can be used.

  2. Evidence that learned generators compose is weak. Although the abstract states that unknown symbols are learned from data, the learned-symbol variant in Table 3 does not fit the seen systems and stays near baseline level on the unseen system.

[1] Zhang, J., Wang, Y., & Lin, G. (2026). LegONet: Plug-and-play structure-preserving neural operator blocks for compositional PDE learning.

[2] Gopakumar, V., Gray, A., Giles, D., Zanisi, L., Kusner, M. J., Betcke, T., ... & Deisenroth, M. P. (2026). Learning Physical Operators using Neural Operators. 

[3] Serrano, L., Han, J., Oyallon, E., Ho, S., & Morel, R. (2026). Test-time generalization for physics through neural operator splitting.

Questions for authors
  1. Could you compare HGL against methods that use the same information (cf. weakness 1), such as a classical solver given the same equation and coefficients and physics-based hybrid methods such as [1], [2], and [3]?

  2. Does the learned symbol result in Table 3 change with longer training and multiple seeds? Plus, can you show composition with learned principal operators in a setting where the mechanisms interact, such as with spatially varying velocity or non-periodic boundaries?

1 comment

W1, Q1: On Fisher–KPP, HGL receives the analytic operators and the true coefficients, so the comparison is not at equal information. A classical solver and hybrid methods [1–3] should be compared.

Thank you for this suggestion. Following it, the revised paper compares HGL with the classical solver and with [1]–[3]. When the governing equation is fully known, HGL coincides with the classical operator-splitting solver. When some mechanisms need to be learned, all methods receive the same known operators and mechanism coefficients. With this equal information, HGL, the physical-operator model, LegONet, and CompNO fit the training trajectories to relative errors below 1%, while HGL generalizes to unseen combinations 15 to 28 times more accurately.

Unseen Fisher–KPP systemstep 10step 80
Classical operator-splitting solver (HGL with ρ_m ≡ 0)7.7×10⁻⁶3.5×10⁻⁵
HGL0.0000660.00031
Best solution operator0.500.47
Five-mechanism, step 60Training systemsUnseenError ratio (baseline / HGL)
HGL (order-constrained)0.00080.0034—
Physical-operator model ([2])0.00130.09528×
LegONet ([1])0.00130.05015×
CompNO0.00590.09026×
Operator splitting ([3])0.00160.07422×
Mechanisms observed only in mixtures
HGL (order-constrained)0.000710.0094—
Generic MLP terms0.00140.1617×

The gap comes from whether each learned term represents its own mechanism. Summing the mechanism terms and applying them one after another by operator splitting give the same result (0.095 in both cases). Training each term to match the time derivative (LegONet) improves the generic MLP terms from 0.095 to 0.050, but they remain 15 times above the order-constrained terms (0.0034). These models share HGL's known operators and the same network per mechanism term, without the order constraint. They differ in training objective, in whether the terms are summed or applied by operator splitting, and in whether the mechanism coefficients are given or estimated from four context frames (Appendix D, Table 8). All baselines receive the same mechanism coefficients as HGL, and in Table 2 they additionally receive three instantaneous known-operator tendencies. A solution operator learns the full flow map and needs to extrapolate for each unseen combination. HGL learns the right-hand side and integrates.

W2, Q2: The evidence that learned generators compose is weak, because the learned-symbol variant of Table 3 does not fit. Does this change with longer training and more seeds? Can composition be shown for interacting mechanisms?

Thank you for raising this point. Learned mechanism terms compose in the five-mechanism experiment that the revised paper sets up following your comment. Transport and diffusion are given as known operators, and three other mechanisms are each learned from single-mechanism trajectories. The test system contains all five. HGL gives 0.0034 at step 60, against 0.095 for the physical-operator model (28 times lower). These mechanisms include those of Fisher–KPP, so the experiment contains Fisher–KPP with a learned reaction (Table 4 of the revised paper). The revised paper also adds an experiment in which the learned mechanisms are never observed alone, and HGL gives 0.0094, against 0.16 for generic MLP terms (17 times lower; Table 5 of the revised paper).

As you asked, we repeated the ablation that removes the order constraint from Table 3 with three seeds and five times the training, and the result is unchanged. In this ablation the analytic form of the transport and diffusion operators is replaced by free per-frequency parameters without the order constraint. The ablation shows that the order constraint is needed: without it, each term can absorb another mechanism's contribution, and the sum remains correct while the individual terms are incorrect. Without the order constraint, the per-frequency parameters are not recovered: they stay within 0.5 of zero, whereas the true values reach 201 and 8×10⁴. On combined-system states, the tendency errors of the generic MLP terms for advection and reaction reach 0.294 and 0.245, while the order-constrained terms stay within 0.009 and 0.011. With order-constrained terms, the learned mechanisms compose at 0.0034, against 0.095 for the physical-operator model (28 times lower).

The mechanisms in these experiments interact. The learned advection velocity depends on the state, so it varies in space and acts together with transport, diffusion and reaction. The steady problem has Dirichlet boundaries and a spatially varying diffusion coefficient.

References: [1] Zhang et al., 2026. [2] Gopakumar et al., 2026. [3] Serrano et al., 2026.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.