acceptodds
Under review as a conference paper at ICLR 2027

Source Anchoring for Physical Consistency in Flow Matching Models

15 views

Abstract

Deep generative models are used to solve partial differential equations and model distributions of physical system states, but ensuring that the generated samples satisfy the governing laws remains challenging. Projection-based flow-matching methods enforce physics by correcting the flow from an unconstrained noise distribution. These corrections shift the generated samples away from the distribution of target solutions, especially in high noise regions. To address this limitation, we propose Source Anchoring for Physical Consistency (SAPC), a Functional Flow Matching method that encodes the physical constraints into the source noise before generation begins. We evaluate SAPC on five systems governed by partial differential equations, covering six tasks with linear and non-linear dynamics, and compare results against five baselines and the unconstrained backbone. Anchoring the source reduces the need for large corrections that drive samples onto admissible but off-distribution states, and SAPC reproduces the target distributions most accurately on every evaluated task, while matching the constraint precision of the best projection-based baselines. Ablation experiments show that this gain arises from pairing source projection with a matched training objective that regresses toward the projected source. These results identify the source distribution as a key design choice for physically consistent generative modelling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Diffusion Model, Physically Constrained Generation, Partial Differential Equations, Conservation Laws

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_qvHuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • The paper introduces Source Anchoring for Physical Consistency (SAPC), a Functional Flow Matching method that projects the source noise onto a constraint manifold before generation. By incorporating this projection into training and additionally projecting the estimated endpoint during sampling, the method aims to improve constraint satisfaction while reducing distributional mismatch.
  • The method is evaluated on six tasks across five PDE systems, comparing against several recent baselines. The results demonstrate substantial improvements in distributional fidelity while maintaining constraint satisfaction comparable to the strongest projection-based methods.

Major strengths/weaknesses and questions are listed below; here are some further minor notes:

  • Related work: Fine-tuning approaches appear to be missing from the discussion, especially regarding the statement "Such approaches are tied to the selected system and must be redesigned for each new problem" (lines 139–140), since fine-tuning exactly tackles this problem of quickly adapting a pre-trained model to various settings. PCFT or PIRF might be suitable references. Otherwise, the related work nicely contextualizes the contribution within recent literature.

  • Physical constraints: The statement in lines 80–84 is somewhat misleading. The considered constraints enforce necessary, but generally insufficient, conditions for solving the PDE, thereby encoding partial information about the dynamics. However, this limitation is presented as a problem of existing methods, even though SAPC retains it.

  • Physical consistency: The definition of the "manifold of physically-consistent states" (Eq. 1, lines 173–187) is confusing. The set contains states satisfying the selected constraints, but not necessarily the governing PDE. This conflicts with the earlier definition of physical consistency through the full PDE residual. "Constraint-admissible states" might be more appropriate. Also, calling this set a manifold requires regularity assumptions.

  • Boundary conditions: The definition of the local constraint RlocR_{\mathrm{loc}} (lines 180–184) is narrower than suggested, as it only directly represents Dirichlet conditions rather than general boundary conditions.

  • Projection formula: The inverse in line 202 does not necessarily exist. The implementation already accounts for this through a pseudoinverse, which could also be reflected in the main text using the Moore–Penrose pseudoinverse.

  • Training and sampling: The wording in lines 207–209 could more clearly distinguish the source projection, performed during both training and sampling, from the posterior-mean projection, which is performed only during sampling.

  • Sampling algorithm: The update in Figure 2 and Algorithm 2 is called an Euler update. While technically defensible, "corrected Euler update" or a reference to the ECI-style update might be more descriptive. Furthermore, the indexing in Algorithm 2 is imprecise: the loop index is not used explicitly, and tit_i would be needed. Replacing AtA_t and BtB_t with their expressions in terms of tt could also make the algorithm more self-contained and easier to read.

Critical strengths
  • S1 Clear motivation and positioning The paper addresses an interesting and relevant issue in physically constrained generative modeling: improving constraint satisfaction without sacrificing distributional fidelity. The distinction between constraint violations and distributional mismatch is compelling, and the paper is generally well motivated and nicely contextualized within recent literature.
  • S2 Extensive and convincing experimental evaluation The empirical evaluation is a major strength. The authors consider several different PDE systems, compare against a broad range of recent baselines, and provide extensive ablation studies. The improvements are substantial and consistent across the considered tasks. In particular, Figure 4 nicely demonstrates the gains in both distributional fidelity and constraint satisfaction.
Critical weaknesses
  • W1 Limited methodological novelty and insight into the proposed mechanism

    • The methodological contribution appears rather incremental. The proposed algorithm essentially modifies an existing Functional Flow Matching approach by projecting the source distribution onto the constraint manifold and adapting the training objective accordingly, while retaining an existing projection-based sampling strategy. While this is a sensible modification, the mechanism itself is straightforward.
    • The theoretical analysis provides some results on constraint residuals along the interpolant, which are a nice addition. However, these are essentially expected interpolation-error bounds for general smooth functions. They establish that interpolation between admissible endpoints preserves affine constraints and approximately preserves nonlinear ones, but provide limited insight into the actual generative mechanism or the improved distributional fidelity.
    • The central question is therefore why source projection leads to such substantial improvements in distributional fidelity. The authors provide an intuitive explanation: starting from a constrained source keeps the generative trajectory closer to the admissible set, reducing the need for large corrective projections that would otherwise distort the generated distribution.
    • However, the empirical evidence does not fully establish this interpretation. In particular, the ablations with endpoint projection but without source anchoring achieve distributional fidelity comparable to unconstrained FFM, which does not directly demonstrate that projection itself introduces additional distributional bias. Furthermore, the conditional FFM baseline achieves competitive distributional metrics without source anchoring. These observations at least partially challenge the proposed explanation and suggest that other effects might play a dominant role.
    • Consequently, I find the explanation of the underlying mechanism somewhat limited. A deeper understanding of when and why source anchoring improves distributional fidelity would significantly strengthen the contribution and be valuable for future developments. Relatedly, some claims regarding "distributional bias" and the mechanism behind its reduction appear stronger than what is directly supported by the experiments or theoretical results. More conservative wording would be appropriate.
  • W2 experimental details I was surprised by the relatively poor performance of PBFM, which I have found to be a competitive baseline in my own experience. After examining the provided source code and comparing it with the original implementation, I identified three potential issues that might contribute:

    • 1) Loss computation There might be an issue in the loss computation for PBFM, at least regarding the heat equation. PBFM uses the callback residual_from_reference to supply the specific per-sample target, but this callback is absent for the heat IC constraint. Therefore, this baseline might use a wrong physics target during training by enforcing a fixed initial condition rather than the actual per-sample constraint. (see src/projectors/constraints/factory.py, lines 54-77)
    • 2) Custom implementation The provided code uses a custom implementation of ConFIG, that at least differs by using torch.linalg.solve (plus fallback) vs. ConFIG's original torch.linalg.lstsq. This might be more of a robustness issue especially for near ill-conditioned systems, but using the official conflictfree package would at least be a simple reproducibility improvement.
    • 3) Methodoogical The original PBFM paper conditions its model on physical parameters (e.g. Reynolds number). Both training and sampling pass these parameters as variables to the network. An exception is the Darcy case, in which the joint distribution over state and parameters is modeled directly. Nonetheless, this means that the original method receives or models the conditioning parameters. For comparison against SAPC, which also does not receive conditioning input, an unconditional PBFM variant might be a meaningful ablation. Generally, however, I believe it would be fair to also inform PBFM about the constraint during inference through its native mechanism instead of withholding conditioning information.

The paper introduces and evaluates a seemingly overlooked mechanism in constrained generative modeling. While the methodological contribution is rather incremental, with a straightforward implementation and limited insight into the workings of the underlying mechanism, the experimental results are convincing and suggest that the proposed approach could be of interest to the ICLR community. I encourage further analysis and discussion of the raised empirical issues as well as a more conservative phrasing around claims interpreting how the method works.

Furthermore, some minor notational/mathematical issues are flagged above. Fixing these would improve readability and accessibility.

Questions for authors
  • Q1 (W1) Could the authors provide further insight into why source anchoring leads to such substantial improvements in distributional fidelity? In particular, how does the proposed explanation of projection-induced distributional bias relate to the ablation results, where endpoint projection alone achieves fidelity comparable to unconstrained FFM, and conditional FFM achieves competitive distributional fidelity without source anchoring?
  • Q2 (W2) Could the authors clarify the potential implementation and experimental issues raised in W2 and verify whether these affect the reported results (also the other experiments) or overall conclusions?
  • Q3 Could the authors compare SAPC against conditional FFM combined with the same sampling-time endpoint projection? Given the competitive distributional fidelity of condFFM, this would help isolate the specific benefit of source anchoring.
5 comments

We thank the reviewer. The revised paper is uploaded with changes in blue.

Q3. We added this comparison (448–459; App. E, Table8, Fig.12), evaluating conditional FFM alone, with the same endpoint projection and sampler as SAPC, and with the full SAPC procedure. Alone, condFFM is competitive in distribution, but its CEL\mathrm{CE}_L is orders of magnitude higher than SAPC’s. With the same endpoint projection, distributional fidelity is comparable, but SAPC has a lower CEL\mathrm{CE}_L on every task, as it needs 1.31.3–3×3\times smaller endpoint corrections. Adding source anchoring to condFFM further lowers its CEL\mathrm{CE}_L on five of six tasks and improves its distribution on Heat and RD. Finally, when the conservation law carries trajectory-specific information (App.E, Table9), SAPC exploits it better than an input channel. The specific benefit of source anchoring is thus tighter constraint satisfaction and smaller corrections without an input channel, and the two mechanisms are complementary.

Q1, W1. To clarify the mechanism, we added ablations anchoring the source on the IC/BC or the global law only (App.D, Table7, Fig.11), a comparison with condFFM (Q3), and a measurement of the endpoint corrections (Fig.12). Anchoring on the IC/BC only reproduces the gain of SAPC, whereas anchoring on the global law only leaves the distribution at the level of S1. The main driver of the distributional gain is therefore the conditional information carried by the anchored source. S1 adds no information and stays close to FFM, and condFFM is competitive because it receives the same information explicitly. Projection-induced shift is only a secondary effect, where corrections are large. We now present source anchoring as a form of conditioning that requires no input channel nor additional sampling cost, complementary to explicit conditioning (18–20, 98–107, 264–269), and revised the claims on distributional bias accordingly. Finally, Eq.4 concerns admissibility along the training interpolant rather than distributional fidelity; we now report when its assumptions hold (Table5, App.A.3).

Q2, W2. We re-ran PBFM in three variants, one for each point, on all six tasks with 3 seeds ×\times 4 test sets. V1) On Heat, the residual_from_reference callback was missing for the IC constraint. We added the per-sample physics target on every benchmark; the effect is small, and the other tasks stay within seed variability. V2) We replaced our implementation with the official conflictfree package. Our implementation treated the complex gradients of the FNO spectral weights as complex vectors, whereas all re-runs now flatten them to real views. Results are unchanged within error on every task. Since V1) and V2) remain within seed variability of the submitted run, the conclusions of the other experiments are unaffected. V3) We now give PBFM the constraint through its native mechanism, conditioning the network on the IC/BC through an input channel during training and sampling. This improves PBFM, reducing PCA-FD by one to two orders of magnitude. CondPBFM still does not enforce the constraints: its CE stays between 0.10 and 0.90, four to six orders of magnitude above SAPC on the affine tasks and RD, and 33 to 8×8\times above it on Stokes. In distributional fidelity, SAPC remains better on Heat, RD and Burgers, while condPBFM is better on Stokes and NS (as condFFM).

Minor points. In the Related Work, we added fine-tuning approaches, citing PCFT and PIRF (140–142). We rewrote the Introduction to state that ICs, BCs and integral conservation laws are necessary but not sufficient conditions for solving the PDE (81–86), and the new Limitations paragraph states that this limitation is shared by every projection-based method, SAPC included (521–525); a new per-sample PDE residual confirms it empirically (App.C.3, G). Eq.1 now defines the set H\mathcal{H} of constraint-admissible states, and we replaced “manifold” with “set” or “admissible set” throughout; regarding regularity, we state after Eq.1 that H\mathcal{H} is a closed affine subspace for linear constraints and a smooth surface for C2C^2, independent non-linear constraints (192–193). After the definition of the local constraint, we state that RlocR_{\text{loc}} represents Dirichlet-type conditions (184). We replaced the inverse with the Moore–Penrose pseudo-inverse in the main text (208) and in App.A.2 and C.1, consistently with the implementation, which uses a Cholesky factorisation and falls back to a pseudo-inverse when AA⊤AA^\top is rank-deficient. We rewrote lines 213–217 to separate the source projection, performed at training and sampling, from the endpoint projection, performed only at sampling. Finally, we call the update a projected Euler update citing ECI (241), corrected the caption of Figure2, and rewrote Algorithm 2 on the explicit grid tit_i, so that every step uses the loop index, with AtA_t and BtB_t written explicitly in terms of tt (244).

The results of the three variants of PBFM and of SAPC are reported below as PCA-FD / CE, mean over 3 seeds × 4 test sets:

  • PBFM (submitted): Heat 496 / 0.65 · RD 630 / 0.69 · Stokes IC 245 / 0.69 · Stokes BC 432 / 0.67 · Burgers 977 / 0.64 · NS 1.1⋅1041.1\cdot10^4 / 0.85
  • V1 (per-sample target): Heat 488 / 0.71 · RD 599 / 0.68 · Stokes IC 245 / 0.66 · Stokes BC 430 / 0.68 · Burgers 997 / 0.71 · NS 1.2⋅1041.2\cdot10^4 / 0.81
  • V2 (official ConFIG): Heat 488 / 0.63 · RD 595 / 0.67 · Stokes IC 244 / 0.68 · Stokes BC 435 / 0.65 · Burgers 981 / 0.62 · NS 1.2⋅1041.2\cdot10^4 / 0.81
  • V3 (conditioned): Heat 2.61 / 0.10 · RD 33 / 0.90 · Stokes IC 23 / 0.16 · Stokes BC 10 / 0.28 · Burgers 2.27 / 0.16 · NS 4.9⋅1034.9\cdot10^3 / 0.25
  • SAPC: Heat 0.48 / 2⋅10−62\cdot10^{-6} · RD 1.01 / 3⋅10−63\cdot10^{-6} · Stokes IC 56 / 0.061 · Stokes BC 56 / 0.036 · Burgers 1.22 / 2⋅10−62\cdot10^{-6} · NS 6.0⋅1036.0\cdot10^3 / 7⋅10−77\cdot10^{-7}

Dear authors, thank you for the revision and the clarifying comments.

My issues regarding presentation and the minor comments are resolved. I believe that the revised paper is easier to follow and particularly appreciate the now consistent mathematical presentation.

Also, W2 seems addressed sufficiently. This resolves the questions regarding the experiments, especially with respect to PBFM.

I am happy to keep my initial positive assessment.

Regarding W1/Q1:

  • I appreciate the more careful wording in the revision. However, some passages still contain strong claims which are not fully supported by the theoretical insights into the mechanism (e.g., lines 94–95).
  • The new ablations (Appendix D, Table 7) and comparisons with conditional FFM (Appendix E, Table 8) provide meaningful insights, suggesting that implicit conditioning, rather than reduced projection-induced bias, is the main driver of the distributional improvements.
  • However, the theoretical explanation remains somewhat limited. In principle, flow matching can transport different source distributions to the same target distribution. While Eq. (4) establishes proximity to the admissible set, it does not explain why source anchoring should improve distributional fidelity.

Overall, the revision strengthens the paper, but I still believe the remaining gap in the theoretical understanding of the central mechanism would be a valuable contribution.

Reviewer_KL4FOpenReview ↗
Rating
3(+1)
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SAPC (source-anchored projection constraints), which applies projection to the source/noise end of the generation process to reduce the distribution shift that projection induces at the terminal end, while enforcing affine and nonlinear constraints across multiple PDE tasks. The method is evaluated on several PDE benchmarks with distribution metrics (W2, FPD), constraint error, multi-seed runs, ablations, and trajectory analysis. The results show large gains on the main tables and clearly separate "low constraint error" from "high distribution quality" as two different objectives. The key claim is that anchoring the source and then projecting provides a benefit beyond conditioning the model and projecting the final output.

Critical strengths
  1. Clear problem and an intuitive, relatively simple method. The source-distribution design has a direct geometric interpretation, and the implementation is simple enough to be adopted.
  2. Broad and well-organized empirical study. It covers multiple PDEs, affine and nonlinear constraints, multiple random seeds, ablations, and trajectory analysis, with significant gains on the main tables.
Critical weaknesses
  1. The conditional-information explanation is not ruled out. Appendix Table 6 shows condFFM already beats SAPC on several distribution metrics (though with higher constraint error). The source projection does both geometric constraint handling and conditional-information injection—the projected noise already contains the specified initial state. The gains may therefore come partly from more effective conditional generation rather than from reducing projection-induced distribution shift. condFFM + the same terminal projection, and a final-output-only projection, are missing.

  2. The "source projection was not explored" claim is too strong. The cited divergence-free diffusion work (Section 4.1) already uses Uσ = U0 + σPZ and initializes from projected noise. SAPC's unified design over FFM, constraint types, and nonlinear projections still has value, but the claim should be narrowed to these specific differences, with direct comparisons.

  3. Constraint satisfaction is partly presented as full physical correctness. H typically encodes only initial values, boundaries, and integral balance, not the full local PDE. For the heat equation, u(x,τ)=sin(x+φ) satisfies the given initial value and periodic boundary and keeps total mass at zero, but does not decay in τ, so it generally does not satisfy the heat equation.

  4. The distribution metrics do not validate the joint space-time distribution. W2 is position-wise averaged and FPD is frame-wise averaged, so they miss cross-position and cross-time dependencies; correct per-position marginals can still be stitched into implausible trajectories. Space-time correlations, energy spectra, and whole-trajectory distances are needed.

  5. The two key controls are missing. There is no "condFFM + projection" comparison and no "source projection with matched training, without terminal projection" comparison, so it is not shown that source anchoring itself has an irreplaceable advantage.

  6. The organization should be improved. Paragraphs are too long, which blurs the logical hierarchy and makes the content harder to follow.

Questions for authors
  1. Can you compare condFFM + the same terminal projection, and a version that only projects the final output, against SAPC under the same training and compute budget? How much of the gain remains attributable to source anchoring itself?
  2. Given that the cited divergence-free diffusion work already uses Uσ = U0 + σPZ and initializes from projected noise, how do you narrow the novelty claim to the specific differences of SAPC (FFM setting, constraint types, nonlinear projections)? Can you provide direct comparisons with that method?
  3. What exactly does H enforce—initial values, boundaries, integral balance, or the full local PDE? Can you provide a test that distinguishes "satisfies the specified constraints" from "satisfies the PDE," e.g., the heat-equation counterexample above?
  4. Can you add space-time correlation functions, energy spectra, whole-trajectory distribution distances, and distribution tests of inferred physical parameters to validate the joint space-time distribution beyond position-wise and frame-wise averages?
  5. Does the theoretical treatment of nonlinear projection hold under the assumptions used in the experiments? Can you clarify the scope of applicability?
2 comments

We thank the reviewer. The revised paper is uploaded with changes in blue.

Q1, W1, W5. We recognise that source anchoring is a conditioning mechanism and revised the paper accordingly (18–20, 98–107, 264–269). We added the experiments required, with the same backbone, projector and training: FFM with only the final output projected; source projection without endpoint projection (S4); conditional FFM in three variants: alone, with SAPC terminal projection and sampler, and with all SAPC projections (432–459, App.D, Table7, Fig. 11, App.E, Tables8–9, Fig.12). Final-output-only projection enforces the constraints but leaves the distribution at the FFM level. S4 matches SAPC's distributional fidelity but leaves the constraints unenforced. condFFM alone is competitive in distribution, but its CEL\mathrm{CE}_L is orders of magnitude higher. With the same terminal projection, fidelity is comparable, but SAPC has a lower CEL\mathrm{CE}_L on every task and needs 1.31.3–3×3\times smaller endpoint corrections. Adding anchoring to condFFM lowers its CEL\mathrm{CE}_L on five of six tasks. Compared with projecting only the endpoint, most of SAPC's distributional gain comes from anchoring the source. When the same information is supplied through an input channel, anchoring's contribution lies in satisfying the constraints more tightly and in reducing the corrections required by the endpoint projection. Anchoring adds value when the constraints carry information that an input channel does not exploit (App.E).

Q2, W2. We agree. The Related Work now acknowledges that Genuist (2026) already project the noise of a score-based diffusion model to enforce fluid incompressibility (167–172). SAPC extends source projection to FFM and to the joint enforcement of local and global constraints, linear and non-linear, and analyses its interaction with endpoint projection and conditioning. We do not re-implement divergence-free diffusion, since all our methods share the same FFM backbone and projector to isolate how the constraint is enforced. The closest analogue in our setting is SAPC without endpoint projection on the linear benchmarks. This variant reaches the same distributional fidelity as SAPC but leaves the constraints largely unenforced, whereas SAPC, with endpoint projection, also enforces them at no distributional cost (Table7).

Q3, W3. We agree. H\mathcal{H} enforces only the IC/BC values and the integral conservation law of each task, not the PDE. We now say so explicitly in (81–86,182–193), and state it as a limitation shared by every projection-based method (521–525). We also replaced "physically-consistent manifold" with "set of constraint-admissible states" throughout. We added a per-sample PDE residual, with an in-range variant that restricts the fit to the physical range of the data (App.C.3, 1365–1378; App.G, 1944–1950; Tables13–14). A sample that satisfies the constraints but does not evolve, as in the reviewer's example, obtains the worst value, rΛ=1r_\Lambda = 1. SAPC has a lower residual than PCFM, ECI and CAFM on every task but a higher one than the unconstrained FFM on RD and Stokes: we conclude that projection enforces the constraints, not the dynamics, which must still be learned from data.

Q4, W4. We added the metrics suggested by the reviewer: space–time autocovariance (ACov), log-spectral distances of the energy spectra in space and time (LSDx\mathrm{LSD}_x, LSDt\mathrm{LSD}_t), whole-trajectory distances (sliced Wasserstein, energy distance, PCA-based Fréchet distance), and the KS distance of inferred physical parameters (312–320, 378–380; App.C.3, 1312–1361; App.G, 1912–1926; Tables13–14). SAPC is the best method on every joint metric and task except LSDx\mathrm{LSD}_x on NS. The margin is one to three orders of magnitude on Heat, RD and Burgers, and smaller on Stokes and NS. SAPC also has the lowest parameter KS on Heat and Burgers.

Q5. Eq.4 assumes that the residual is C2C^2 with bounded Hessian and both endpoints of the training interpolant lie on H\mathcal{H}. The first holds for all our constraints, as the IC/BC pins are affine and the non-linear conservation laws are polynomial in the state. The second holds as far as the source projection converges (Table5): exactly on the affine tasks, up to 22–6×10−66\times10^{-6} on RD and Stokes BC, while on Stokes IC the training source stalls at about 0.670.67, where the bound holds only with an additional term t ∥R(Vˉ1)∥t\,\|\mathcal{R}(\bar{\mathcal{V}}_1)\| (App.A.3). The bound covers the training interpolant, not the sampling trajectory, which is corrected at every step; empirically, the sampling residual follows the predicted profile on RD but not on Stokes. Since the Gauss–Newton projector converges only locally, we report convergence diagnostics rather than global guarantees (App.A.3, 770–809; App.C.1, 1158–1171; Sec. 4, 461–463, 513–517).

W6. We split the main text into shorter paragraphs and introduced each result with a bold heading.

Q5 (scope of applicability). The scope of applicability deserves a dedicated discussion.

The bound applies to any affine constraint, and to non-linear constraints that are C2C^2 near H\mathcal{H}, provided the source projection converges. Beyond that, it weakens progressively, with an additional term proportional to the residual of the projected source (see App.A.3). Two limits follow. First, the bound concerns the training interpolant, not the sampling trajectory, which is corrected by the endpoint projection at every step and by a final full projection at t=0t=0. Empirically, we show that the sampling residual follows the predicted profile on RD, peaking near t=1/2t=1/2 and vanishing at both ends, but not on Stokes, where the training source has not fully converged (Sec. 4, 513–517; App. G). Second, the projector is a local Gauss–Newton approximation of the metric projection onto a curved set, as in PCFM, so its convergence is guaranteed only near H\mathcal{H}.

For a new problem, the applicability of SAPC can be checked a priori by verifying that the constraints are affine or smooth, and a posteriori by monitoring the residual of the projected source, as in Table 5. When the source projection converges, the theory holds; when it does not, as on Stokes IC, the residual quantifies how far the assumptions are violated.

Reviewer_Dx8HOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces Source Anchoring for Physical Consistency (SAPC), a flow-matching framework that tries to encode physical constraints directly into the source noise by projecting it. The projection of the noise happens during both training and inference. The authors showcase results on six tasks across five different PDE systems, covering both linear and nonlinear constraints.

Critical strengths

S1. The method seems well-motivated, is clearly demonstrated, and is easy to understand and follow. The authors also provide some ablations to clarify the method and its contributions.

S2. The writing is clear and easy to follow.

Critical weaknesses

W1. One of my main concerns is that the paper does not properly distinguish the type of generative modeling task considered by the existing baselines. Typically, in methods such as ECI and PCFM, the flow-matching model is trained on a distribution with varying physical parameters, ICs, and BCs. During sampling, only a specific set of constraints is imposed, without explicitly retraining the flow-matching model. This is an important aspect of these methods, as they aim to improve the distributional accuracy and physical consistency of a pretrained generative model.

However, upon closer inspection, the proposed approach seems to be another way of explicitly conditioning the flow-matching model during training.

Consider the simple case of linear constraints, where (AV=b). Projecting Gaussian noise (Z) gives

Pb(Z)=Z−A⊤(AA⊤)−1(AZ−b).P_b(Z)=Z-A^\top(AA^\top)^{-1}(AZ-b).

The training interpolation is then

Vt=(1−t)V0+tPb(Z).V_t=(1-t)V_0+tP_b(Z).

Since both the target (V0)(V_0) and the projected source satisfy AV=bAV=b, we have AVt=bAV_t=b at every timestep. In the case of IC/BC constraints, this means that the model already receives the specified IC/BC information throughout training, including at the highest noise levels. This is effectively a way of conditioning the flow-matching model during training, rather than explicitly concatenating the IC/BC as additional input channels.

This interpretation is also supported by the appendix results. In Table 6, conditional FFM (condFFM) is often quite close to SAPC and even performs better on the Wasserstein distance in four of the six tasks. Importantly, this is without applying the same inference-time projection used in SAPC.

In hindsight, SAPC seems closely related to conditioning the flow-matching model during training, followed by projection-based sampling at inference. While the two approaches are not mathematically identical, the paper does not clearly establish what additional benefit comes from source anchoring beyond this conditioning effect. I think this substantially weakens the novelty claim and makes the comparisons with the existing inference-time constrained generation methods difficult to interpret.

W2. Another key issue is that the paper claims that SAPC has comparable inference cost to the existing baselines. However, as shown in Table 5, SAPC is approximately 2-7 times slower on all tasks other than Navier-Stokes. I don't think the claim of comparable inference performance is justified across the demonstrated benchmarks. Moreover, there's no discussion of why SAPC is particularly faster on Navier-Stokes.

W3. I don't explicitly understand why the projection of a Gaussian source noise onto the manifold of the constraints that are defined for the PDE solution will yield any significant gains. The proposed methods use Gauss-Newton-style updates, which rely on local linearization of the constraints. The initial Gaussian noise is not meant to satisfy these constraints and may be far from the constraint manifold, where the convergence of the projection is not necessarily guaranteed and becomes a bigger problem with nonlinear constraints. Moreover, the paper does not clearly demonstrate whether these source projections converge, especially when they are specifically doing only a fixed number of iterations.

W4. While I encourage the ablations provided in the appendix, they do not sufficiently isolate the advantage of source anchoring over explicit conditioning (as I briefly discussed in W1). I feel the most important missing comparison is the condition FFM with the same inference time projection and the sampling procedure as SAPC. As discussed in W1, the condFFM method already achieves comparable or even better distribution accuracy. It is therefore unclear to me whether SAPC would provide any meaningful improvement over a conditional flow matching model with the same inference procedure. It seems like SAPC is one of the "implicit ways" to do conditional flow matching on ICs and BCs. Correct me if I am wrong.

Questions for authors

See the weaknesses above, but here are some additional questions:

Q1. The test results seem to use the same parameter ranges and IC/BC families used during the training. How well does the method generalize to some out-of-distribution tasks, such as unseen parameters or different IC and BC cases?

Q2. Upon close inspection, Figures 3 and 11 show close agreement with the ground-truth ensemble statistics, which may seem like a sign of overfitting. Can the authors provide an additional evaluation of individually generated trajectories (e.g., PDE residuals or inferential gradient residuals) to demonstrate that the model captures the underlying distribution rather than merely matching pointwise statistics?

Q3. Can the authors shed some light on the convergence diagnostics for anchoring source samples? I don't understand why Gaussian would satisfy constraints originally defined for PDEs.

Q4. Another interesting ablation would be to construct the source projection operator using only the IC/BC constraints, without enforcing the additional global conservation laws. This would help clarify whether the reported improvements primarily come from implicitly conditioning the flow-matching model on the IC/BC information, as discussed in W1, or whether enforcing the global conservation laws during source projection provides additional benefits in distributional accuracy.

3 comments

We thank the reviewer. The revised paper is uploaded, with changes in blue.

W1,W4 We agree, and the paper now presents source anchoring as a form of conditioning that needs no input channel, with the contributions revised accordingly (18–20; 98–107; 264–269). We also clarify that SAPC is not zero-shot: anchoring only at sampling on an unconstrained backbone (ablation S3) leaves the constraint error at the FFM level (435–442, Table7). All baselines share the same backbone, trained on the same data with varying IC/BC (1185–1187). To isolate what anchoring adds over explicit conditioning, we compare SAPC with variants of the conditional FFM: alone, with the same endpoint projection and sampler, and with the full SAPC projections (448–459; App.E, Table8, Fig12). Alone, condFFM is competitive in distribution, but its CEL\mathrm{CE}_L is 66–1313 orders of magnitude higher. With the same endpoint projection, distributional fidelity is comparable, but SAPC has a lower CEL\mathrm{CE}_L on every task (by 2×2\times to 106×10^{6}\times) and needs 1.31.3–3×3\times smaller endpoint corrections; adding anchoring to condFFM further lowers its CEL\mathrm{CE}_L on five of six tasks. The two mechanisms are therefore complementary rather than interchangeable.

W2 We removed the claim of comparable cost. Table 6 now reports time per sample and projector time (464–468; App.C.4, 1396–1413). Anchoring adds one projector call per sample, so the cost is set by the endpoint projector. With affine constraints, SAPC costs 1.031.03–1.11×1.11\times FFM on Heat and Burgers, 1.71.7–1.9×1.9\times less than PCFM and CAFM, and is about 30×30\times faster on NS, where PCFM and CAFM correct one sample at a time. With non-linear constraints, Gauss–Newton dominates and SAPC is 1.61.6–7.3×7.3\times slower than PCFM and CAFM. ECI is 4.94.9–6.8×6.8\times slower than SAPC on every task.

W3,Q3 The constraints act on the state, therefore apply to any field, Gaussian noise included. When anchoring, the projection overwrites the noise on the IC/BC region with the prescribed values and corrects it along the conservation integral. This is a form of conditioning (W1). Table 5 reports the residual ∥R∥∞\|\mathcal{R}\|_\infty of the projected source against the number of iterations kk (461–463; App.C.1, 1158–1171). The projection is exact in one step on the affine tasks, and Gauss–Newton converges from white noise on RD and Stokes BC (to 22–6×10−66\times10^{-6}). On Stokes IC it reaches 10−310^{-3} at sampling but stalls at ≈0.67\approx 0.67 in training, where each draw is re-targeted to a different IC and the weakly damped iteration converges only locally; the training source is then only approximately anchored, and Eq.4 holds up to an extra term (App.A.3).

Q1 Without retraining, we tested IC/BC families absent from training and physical parameters outside the training range, for SAPC, condFFM with endpoint projection, S1, PCFM, ECI, CAFM and FFM (470–476; App.F, Tables10–12). SAPC still enforces the new conditions (CEL<3.2×10−8\mathrm{CE}_L < 3.2\times10^{-8}; CEG<1.8×10−5\mathrm{CE}_G < 1.8\times10^{-5} except on Stokes BC) and has the lowest PCA-FD among unconditioned methods on four of six tasks, but condFFM generalises better on four of six, and no constrained method recovers the target on Heat and Stokes BC. No method extrapolates the free parameter (KS\mathrm{KS} 0.400.40–11 vs 0.030.03 for the reference); SAPC and condFFM keep the lowest SMSE on Heat, RD and Burgers.

Q2 All statistics are computed on four disjoint held-out test sets. We added joint space–time and whole-trajectory metrics, the KS\mathrm{KS} distance of inferred physical parameters, and a per-sample PDE residual with fitted coefficient (312–320, 378–380; App.C.3; Tables13–14). SAPC is best on every joint metric except LSDx\mathrm{LSD}_x on NS, by 11–33 orders of magnitude on Heat, RD and Burgers, and has the lowest KS\mathrm{KS} on Heat and Burgers. Its PDE residual is lower than PCFM, ECI and CAFM on every task, but higher than FFM on RD and Stokes: projection enforces the constraints, not the dynamics, as now stated in the Limitations (521–525).

Q4 We added both ablations (App.D, 1495–1508; Table7, Fig.11). Anchoring on the IC/BC only is comparable to full SAPC on every task; anchoring on the global law only reaches SAPC's constraint error but leaves the distribution at the level of S1. This confirms the reviewer's hypothesis: the gain comes from the IC/BC anchor. We notice that our global laws, however, carry no trajectory-specific information (their right-hand side is zero or computed from the generated field). We therefore tested a new RD variant with a prescribed net boundary flux q=gL−gRq = g_L - g_R (App.E, 1653–1737, Table9). SAPC, receiving qq only through the projection, reproduces the target better than condFFM receiving it as an input channel (PCA-FD\mathrm{PCA\text{-}FD} 0.49±0.100.49 \pm 0.10 vs 1.03±0.331.03 \pm 0.33), with both enforcing qq to 3×10−73\times10^{-7}. This completes the comparison of W1,W4.

I thank the authors for providing such a detailed response and including additional experiments. The paper has been significantly improved by the additional experiments (including comparisons with conditional FFM, additional ablations, and a convergence analysis of the source projection). Most of my questions have been answered in the revised manuscript.

However, I still have concerns about the method's novelty and impact, which I list below.

The authors now recognize that source anchoring corresponds to conditioning during training (this was my main concern in W1). This is an important point, as methods such as ECI and PCFM aim to enforce constraints at inference without retraining the pretrained model. In contrast, SAPC conditions the model throughout training. Therefore, it is not clear that the orders-of-magnitude improvements over these methods demonstrate the benefit of the proposed source-anchoring mechanism.

Most importantly, the additional results in Table 8 show that, with the same endpoint projection and sampling procedure applied to conditional FFM, SAPC no longer yields substantial improvements in distributional accuracy. Heat and RD are improved with SAPC; however, conditional FFM with projection achieves a lower Wasserstein distance on the remaining four tasks. The differences in distributional accuracy are quite small compared to the original reported improvements. Whilst SAPC achieves lower local constraint errors and smaller endpoint corrections, both methods already enforce the local constraints to very low absolute residuals.

I found the new ablation study (Table 7) very informative. Anchoring the source using only the IC/BC constraints achieves almost identical distributional accuracy to full SAPC. However, anchoring using only the global conservation constraints does not produce comparable distributional improvements. This suggests that my original concern was valid and that the improvement primarily comes from incorporating the required IC/BC information into the source during training.

Regarding W3, thanks for providing the additional convergence analysis. However, I can see from Table 5 that the source projection on Stokes IC does not converge during training, with a residual of around 0.67 even after 20 iterations. Despite this, SAPC still provides competitive distributional accuracy. This again raises the question of whether the improvement actually comes from satisfying all the constraints on the source.

The new experiment on the prescribed-flux RD task shows that source anchoring can be beneficial when the global constraint provides additional information. However, this is still another way to incorporate conditioning into the model. The improvement over explicit conditioning is relatively modest and is demonstrated in only one additional setting.

Overall, whilst the experiments have provided much more clarity on how the method works, they also substantially narrow the paper's contribution. It seems that this method is implicitly performing conditional flow matching by projecting the source noise and then using an existing projection-based sampling method. Whilst this could be considered a useful design choice, I do not feel it is a sufficiently novel methodological contribution.

I thank the authors for their efforts and improvements during the rebuttal. I may consider adjusting my original score following the reviewer–AC discussion. However, I still feel that this paper falls below the acceptance threshold for ICLR due to its limited methodological novelty and the lack of a substantial or consistent advantage over the conditional FFM baseline.

We thank the reviewer for reading our revision and acknowledging the improvements.

We agree that source anchoring acts as implicit conditioning, which explains the similar distributional accuracy in Table 8. Beyond this equivalence, SAPC offers two advantages over conditioning through an input channel.

1) Conditional FFM requires an input representation of the prescribed values bb. This is straightforward for IC/BC on a grid, but less direct for constraints that are non-local, integral or non-linear, or whose values are defined on a different discretisation. SAPC uses the constraint residual RR and its projector, which enforces the constraint and also delivers the conditioning, without any additional input channel.

The prescribed-flux RD variant (Table 9) illustrates this distinction. In this experiment, the net flux qq specifies an integral relation among the state values. An input channel supplies qq and leaves the network to learn that relation; source anchoring instead embeds it directly in the source, and hence in every state along the flow. This is an example of a constraint that is natural to anchor but less obvious to encode as an input channel. We agree that this is a single setting and we present it as a proof of concept of this potential; extending it to other constraints that are difficult to encode as inputs is a natural next step.

2) The paper shows that SAPC, by construction, keeps its trajectories close to the admissible set. Conditional FFM learns the conditioning from data and offers no such guarantee, so its trajectories can be far from the set. In both cases, the endpoint projection corrects this drift, so that SAPC and conditional FFM reach similarly low residuals. The difference lies in how much correction is needed along the flow, and how much the endpoint projection must change the network’s predictions. SAPC requires corrections smaller by a factor of 1.3–3 than conditional FFM (Figure 12). This becomes relevant for non-linear constraints, whose projection is solved iteratively and therefore benefits from starting closer to the solution. Figure 12 suggests this effect, which can be explored further in future work.

Source anchoring and explicit input conditioning can also be combined: adding source anchoring to conditional FFM further reduces the corrections along the flow, and lowers the local constraint error on five of six tasks, as well as the global constraint error on Stokes (Table 8). Since conditional FFM already receives bb, anchoring adds no new information, and these improvements can only come from keeping the flow on the admissible set.

We hope that this clarifies the scope of our contribution: a constraint-driven mechanism that both supplies conditioning without a separate input encoding and keeps the training states on or near the admissible set, reducing the correction required during sampling.

Comments on the paper

1 comment

Dear Reviewers,

Please review the authors' rebuttal and participate in the discussion to address any remaining questions before the period closes on Nov 18 AoE.

Thank you!

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.