acceptodds
Under review as a conference paper at ICLR 2027

Physics-Aligned Finite Maps: Recasting Physics-Constrained Generation as Next-State Prediction

17 views

Abstract

Physics-constrained generative modeling aims to generate realistic physical fields that are both distributionally accurate and physically consistent, with applications to inverse problems, conditional generation, and scientific simulation. Yet these two requirements can impose competing objectives: existing methods typically balance generative and physical losses during training or correct samples at inference, leading to trade-offs among fidelity, consistency, and computational cost. We introduce Physics-Aligned Finite Maps (PAFM), which instead factorizes target construction into distributional progression in free coordinates and physics-based completion. At every step, the sample is advanced toward the data distribution in free physical coordinates, while known physics completes this state into a feasible full field. The two objectives are therefore compatible at each teacher next-state, which serves as both the endpoint of the current transition and the starting point of the next. Physics-constrained generation is thus recast as a sequence of finite next-state predictions with sample-wise compatible teacher targets. We instantiate the prior-to-data sequence with a coarse-to-fine heat bridge, and further distill offline physical completion into a lightweight transition recovery network to suppress rollout drift without physical solves between learned transitions. Across three benchmarks, PAFM achieves low physical residuals, competitive distributional fidelity, and millisecond-scale inference, providing an effective framework for physics-constrained scientific generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generation Model, Physics, Scientific Machine Learning, PDE

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_KoXnOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces Physics-Aligned Finite Maps, a novel method for injecting physics constraints into generative models. The method contains two important parts: a physics-aligned teacher and sequence and transition recovery distillation. The physics-aligned teacher sequence model directly trains a generative model in a modified space where each intermediate state satisfies the physical constraint. And the transition recovery distillation models gradually correct errors in the autoregressive rollout predictions of the generative models. The method itself is very interesting and novel, but many unclear aspects of the paper's presentation, along with some weak results, have led reviewers to question its actual effectiveness.

Critical strengths

The method itself is novel. Although the current status of the paper makes it difficult to accurately evaluate the actual merits and drawbacks of this method, the idea itself is quite interesting.

Critical weaknesses

The presentation of the manuscript is poor, with many unclear and confusing concepts and discussions. The most critical issue comes from the definition of "free variables zz", "physical completion map DcD_c", and "free coordinates extractor EcE_c" . It is very difficult to properly understand their actual meaning before getting into the experiment setups. Even with the experiment sections, the reviewer needs to carefully dig into the Appendix to fully locate the corresponding operations in each setup. It will be much better if the author can show some examples of these variables/operations before the method discussion. Meanwhile, the unclear presentation induces many questions about the method itself, which further reduces the reliability of the current work.

Questions for authors
  • What is the "free-coordinate extractor EcE_c" for Kolmogorov flow? Do authors simply use z=xz=x for the Kolmogorov flow?
  • In Eq.(4), since z(tk)z(t_k) is a variable that contains noise, is there any guarantee that the completion map DcD_c is still applicable to noisy inputs?
  • Given a data sample xx and a physics constraint operator R(c)R(c), can the author give some general comment on how to build/choose zz, EcE_c, and DcD_c?
  • From lines 2205 to 225, the author argues that "Crucially, the physical term does not introduce a separate transition destination" since "both terms attain zero at the same prediction". However, this holds for almost all physics-constrained training problems, i.e., if the model accurately predicts the target, the physical residual of the target is zero. What specific point does the author want to emphasize here?
  • In Eq. 10, why is yky_k necessary for the input of the correction network? It seems that yˉk+1,tk\bar{y}_{k+1},t_k and cc are enough for the network to predict the discrepancy.
  • What does the red curve "No-TRD" mean in Figure 1? Since two blue lines represent "BASE+TRD" and "BASE" respectively, does the red line actually mean "No-BASE"?
  • Why do SMSE and WD show different trends in Table 2? For example, PBFM is better than PAFM in SMSE, while PAFM is better than PBFM in WD?
  • Why is there unusual performance decay with more inference steps for PAFM, PCFM, and D-Flow in Figure 2?
  • With fewer inference steps, e.g., N<=5, the performance of PCFM seems to be much better than PAFM. I.e., the PCFM shows a much lower physics residual with slightly higher WD. Can the author give some comments on this?
  • Why is PCFM missing in the experiment of Darcy Flow?

Other comments:

  • It seems that all the test cases are from the PBFM paper. It is better to cite it properly in Sec. 4.1.
5 comments
AuthorsOpenReview ↗

[1/3] Response to Reviewer KoXn

We thank the reviewer for the constructive questions. We have reorganized the definitions and method presentation in the revised manuscript, and all references below follow its numbering.

W1: Definitions of free coordinates, extraction, and physical completion.

We have added a dedicated Preliminaries section (§3) to introduce the physical constraints, free coordinates, and physical completion before presenting the method. §3.2 explains these concepts through definitions and a concrete example. For complete state xx, z=Ec(x)z=E_c(x) contains the quantities used to construct the generative path; DcD_c reconstructs the complete state using the selected physical relations. Both maps are prescribed, not learned. Under F=maF=ma with known m>0m>0 as condition, Eq. (3) gives z=Fz=F, Em(F,a)=FE_m(F,a)=F, and Dm(F)=(F,F/m)D_m(F)=(F,F/m). Table 4 and Appendix C collect the benchmark-specific constructions.

Q1: Is the Kolmogorov extractor the identity?

Yes. For physical constraints that can be enforced by projection, we can choose the identity extractor, Ec(x)=xE_c(x)=x, and use the corresponding projection as the physical completion map. For Kolmogorov Flow, Dc(z)=PLzD_c(z)=P_{\mathrm L}z, where PLP_{\mathrm L} is the Leray projector (Appendix C.1, Eq. (36)). Both velocity channels are retained as a redundant representation. “Free” describes the coordinates used to construct the path; it does not require a minimal set of independent variables. We have clarified this point in the revised §3.2.

Q2: Is completion applicable to noisy inputs?

The presence of noise does not itself invalidate physical completion; what matters is whether the intermediate coordinates remain within the admissible domain of DcD_c. In PAFM, randomness is introduced in appropriately chosen generating coordinates, and the coupled fields are then recomputed from those coordinates. We do not independently perturb an already completed field and assume that it remains physical. Our benchmark constructions address this requirement explicitly (Appendix C.1–C.3). For Kolmogorov Flow, the Leray projection is defined for arbitrary discrete velocity proposals, including noisy ones. For Dynamic Stall, temperature and density are represented in logarithmic coordinates; exponentiation with the specified range control preserves positivity, while the prescribed gradient parameterization permits analytical completion of pressure and wall shear. For Darcy Flow, we perturb log-permeability rather than permeability itself, so K=exp⁡(z)>0K=\exp(z)>0; pressure is then obtained under the fixed compatible forcing, boundary conditions, and zero-mean gauge, subject to numerical solver accuracy. Thus, intermediate states need not resemble clean data to admit physical completion: they must satisfy the completion map’s input-domain requirements. This is why PAFM constructs the stochastic bridge in generating coordinates before completing each teacher state.

Q3: How should zz, EcE_c, and DcD_c be chosen?

Our guiding principle is to identify the free variables according to the known physical relations, rather than arbitrarily reducing the state dimension. For a state x∈Rnx\in\mathbb{R}^n with rr independent physical constraints, the feasible states generally have n−rn-r local degrees of freedom. We preferentially choose the physically determining quantities as zz, define EcE_c to extract them from xx, and construct DcD_c to recover the complete state using the governing relations.

For example, consider Newton's second law, F=maF=ma, with mass mm given as a condition. For the complete state x=(F,a)x=(F,a), the applied force determines the acceleration. We therefore choose z=Fz=F, Ec(F,a)=FE_c(F,a)=F, and Dc(F)=(F,F/m)D_c(F)=(F,F/m). Thus, the generating variables follow the physical dependency structure, while completion reconstructs the remaining quantities.

This decomposition is not necessarily unique; we favor physically meaningful variables and computationally tractable completion operators. When an explicit free/dependent-variable decomposition is unnecessary, as in Kolmogorov Flow, we instead use a redundant representation with projection-based completion. We have clarified these general design principles and examples in the revised §3.2.

AuthorsOpenReview ↗

[2/3] Response to Reviewer KoXn

Q4: What do lines 220–225 mean?

We thank the reviewer for this question. Our key distinction is not merely that both losses share a feasible target, but that each finite next-state prediction directly produces a complete physical field, allowing physical constraints to be imposed without propagating generation to the final state.

In conventional FM-based methods such as PBFM, the network predicts an instantaneous velocity, requiring forward integration to obtain a terminal state for physical supervision. PAFM instead directly learns finite transitions between physically completed teacher states. Each predicted next state is already a complete physical field, so both transition matching and physical residuals can be evaluated immediately, without propagating through the remaining generative trajectory. This avoids costly terminal rollout and its associated backpropagation while aligning training with the updates actually executed during sampling.

This formulation also naturally accommodates different prescribed generation budgets. Increasing the number of finite transitions NN provides a finer progression toward the target distribution, which can mitigate conditional averaging and improve distributional fidelity, as supported by our additional diversity experiments and the analysis in Appendix B.2. TRD further mitigates physical inconsistency accumulated during these transitions.

Thus, the contribution is not simply a shared feasible optimum, but a constrained generative formulation in which every finite update is both a complete physical-field prediction and a directly supervised transport step.

Q5: In Eq. 10, why is yky_k necessary for the input of the correction network?

Yes, the provisional state, together with the transition time tkt_k and condition cc, is sufficient to determine the recovery target. Our input choice in Eq. (12) provides a direct connection between BASE and TRD: we pass BASE’s existing inputs (sk,tk,c)(s_k,t_k,c) together with its output, the predicted increment fθf_{\theta}, to the recovery network. TRD can internally represent the addition s^k+1=sk+fθ\hat{s}_{k+1}=s_k+f_{\theta} and learn the corresponding correction defined in Eqs. (10)–(11), without requiring the provisional state to be separately constructed as its input. This choice therefore reflects how the two networks are connected, rather than a requirement for information beyond the complete provisional state. The revised §4.4 clarifies this relationship immediately below Eq. (12).

Q6: What does the red curve "No-TRD" mean in Figure 1...

We have clarified the three curves in the revised legend and caption of Figure 2 (§5.2.1):

  • BASE-only (red) follows an independent uncorrected rollout BASE.
  • PAFM before TRD (purple) shows the BASE-predicted next state from each recovered PAFM state.
  • PAFM (blue) shows the corresponding state after TRD.

Q7: Why do SMSE and WD show different trends in Table 2?

WD and SMSE capture different aspects of distributional fidelity. WD compares one-dimensional marginal distributions, whereas SMSE measures errors in conditional standard-deviation fields. At N=20N=20 on Dynamic Stall, PAFM improves WD over PBFM (5.59435.5943 versus 7.06187.0618) but has worse SMSE (1.45131.4513 versus 0.87810.8781), using Table 25's scaling. The revised §5.2.2 explicitly acknowledges this limitation. The added diversity analysis (Appendix K.1, Table 27) confirms underestimated conditional variability: PAFM/PBFM variance ratios are 0.6716/0.92530.6716/0.9253, with coverage 83.92%/84.61%83.92\%/84.61\%. PAFM improves with additional transitions, from variance ratio 0.01370.0137 and coverage 34.04%34.04\% at N=1N=1. Appendix A.2, particularly Eq. (22), explains the conditional covariance lost through deterministic regression. Better marginal matching therefore does not establish better conditional spread, and learned physical recovery does not guarantee restoration of this missing variability.

Q8: Why is there unusual performance decay with more inference steps for PAFM, PCFM, and D-Flow in Figure 2?

The observed non-monotonicity depends on the metric and method. For PAFM on Dynamic Stall, WD, JS, and SMSE all improve across N=1,5,10,20N=1,5,10,20 (Figure 3 and Table 2). Only the physical residual increases locally from N=5N=5 to 1010, before reaching its lowest value at N=20N=20. The revised §5.2.2 explicitly distinguishes these trends: more transitions improve ensemble metrics, while physical residuals need not decrease monotonically.

AuthorsOpenReview ↗

[3/3] Response to Reviewer KoXn

For PAFM, finer transitions reduce local averaging effects, while the final rollout also depends on learned-map errors and their propagation. More sampling steps shorten each transition, but the final physical residual also depends on errors accumulated through preceding predictions. Our Darcy diagnostic illustrates this distinction (Appendix I.2, Table 21). As NN increases from 5 to 20, the physical correction required after a BASE prediction decreases from 0.0371 to 0.0285 when BASE receives a teacher state. When it receives states from its own rollout, however, the required correction remains nearly unchanged (0.0383 to 0.0389). Thus, the benefit of shorter transitions observed on teacher states can be offset by accumulated prediction errors during generation. This explains why finer transitions do not necessarily yield monotonically lower final residuals and motivates TRD’s complementary role in correcting the states passed to subsequent transitions.

PCFM and D-Flow exhibit different budget-dependent trade-offs. Appendix K, Table 25, shows that D-Flow’s RE improves from N=10N=10 to 2020 while its WD worsens; PCFM maintains a very low residual while its WD increases beyond N=5N=5. Their base budgets also affect additional inference operations: PCFM performs more physical corrections and projections, while D-Flow optimizes the source through the discretized generator. These mechanisms differ from PAFM’s learned finite transitions, and increasing their base budget does not by itself ensure better distributional fidelity. The results therefore support evaluating physical consistency, distributional quality, and total inference cost together when comparing sampling budgets.

Q9: With fewer inference steps, e.g., N<=5, the performance of PCFM seems to be much better than PAFM...

PCFM is a strong constraint-enforcement baseline: it has much lower Dynamic Stall residuals at all reported budgets and better JS/SMSE at N=5N=5. PAFM retains lower WD at N=1/5N=1/5. The cost difference is substantial: Table 25 reports 20 relaxed physics updates per outer step for PCFM. At N=1/5N=1/5, it uses 22/10622/106 constraint evaluations and 2/62/6 projections, taking 3.6924/16.41123.6924/16.4112 ms per sample; PAFM takes 0.4644/1.70510.4644/1.7051 ms, approximately 8.0×/9.6× faster. Base budget therefore does not imply equal computation. §5.2.2 and Figure 3 make this accuracy–latency trade-off explicit. We do not claim learned recovery matches exact enforcement in residual accuracy.

Q10: Why is PCFM missing in the experiment of Darcy Flow?

Our initial comparison did not include PCFM because evaluating the released sampler in our Darcy setting required a task-specific adaptation to the 64×6464\times64 continuous-permeability fields and Neumann boundary conditions. We have now completed this adaptation and added the comparison, with N=20N=20 sampling steps and a fixed set of 1,024 Gaussian initializations. The adapted numerical projection uses uniform Frobenius scaling of the Jacobian and residual with relative damping 10−610^{-6}.

The results are reported below using the main table’s metrics; all errors are unscaled.

MethodRE ↓WDpWD_p ↓WDKWD_K ↓Valid ↑IT (ms/sample) ↓
PCFM (adapted)0.1882120.0062551.16475977.34%2939.94

RE is evaluated on all 1,024 samples. WD follows the original whole-field positive-permeability filter, retaining 792 samples; Valid reports this fraction. This additional experiment extends the Darcy comparison to PCFM under the stated adaptation and evaluation protocol. Compared with PCFM (adapted), PAFM achieves 78.8% lower RE, 60.4% lower WDp\mathrm{WD}_p, and 88.7% lower WDK\mathrm{WD}_K, while increasing validity from 77.34% to 100% (Table 3). The adapted PCFM sampler also requires a median 2.94 seconds per sample, incurring substantial sampling cost. These results show that, under the evaluated Darcy setting and adaptation, PAFM achieves stronger physical consistency and better marginal fidelity.

Other comment: Attribution of the benchmark datasets.

Thank you for this correction. The revised experimental setup, now §5.1, explicitly attributes Kolmogorov Flow and Dynamic Stall to the PBFM release (Baldan et al., 2026), and Darcy to PIDM (Bastek et al., 2025; Bastek, 2024). Appendices C.1–C.3 document provenance, numerical generation, preprocessing, and splits. PBFM is therefore credited both as a baseline and as the source of the two datasets we reuse.

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

This paper addresses the issue that generated simulation results tend to violate physical laws by employing physics-based computation to produce correct demonstrations, which are then used to teach a network to perform step-by-step generation and error correction. Experiments demonstrate that this approach can complete network generation within a few ms while substantially reducing physical errors.

Critical strengths
  1. The physical constraints are replaced by a small network, which both reduces computational cost and improves accuracy.
  2. It outperforms existing generative methods.
  3. The mathematical derivation is relatively rigorous and detailed.
Critical weaknesses
  1. Compared with the baseline, the improvement on some metrics is relatively limited; although it increases computational cost, it does not achieve the best performance on all metrics.
  2. The proposed approach lacks a clear innovation or extended research direction, as incorporating a network to learn physical constraints is a common practice in many fields.
Questions for authors
  1. Since the entire state can be derived from a subset of physical quantities, would it not be better to directly generate this subset of physical quantities and then compute the remaining quantities in a final step? Why generate output quantities determined by free variables? In your tables, the runtime difference between these schemes is not substantial; a few milliseconds are not critical for practical use. Pursuing a smaller network with better performance seems more reasonable.
  2. In the two rounds of TRD training for Kolmogorov Flow and Dynamic Stall, BASE remains frozen throughout, and a two-stage scheme is used to achieve adaptation. However, can it be guaranteed that after this fine-tuning, BASE will not encounter problems when interacting with the fine-tuned network? Why not consider reinforcement learning or imitation learning approaches that complete the process within a single training stage?
5 comments
AuthorsOpenReview ↗

[1/3] Response to Reviewer hqpE

First of all, thank you very much for your opinions and suggestions. We will explain them one by one. You can also view the revised version, where we have marked the questions you raised to help you quickly locate them.

W1: Limited improvements and additional cost.

On Kolmogorov Flow, both BASE and full PAFM outperform every evaluated external baseline in WD and JS, including FM-OT, PBFM, PCFM, D-Flow, and DiffusionPDE (Table 1). BASE achieves the lowest WD and JS overall, while full PAFM retains nearly identical distributional fidelity and achieves the second-lowest physical residual, surpassed only by PCFM with online projection. Recovery adds just 0.10 ms/sample. These results demonstrate that the finite-transition formulation itself delivers leading distributional accuracy, while learned recovery substantially strengthens physical consistency at negligible additional cost.

On Dynamic Stall and Darcy Flow, PAFM occupies the Pareto frontier when physical residual, distributional fidelity, and inference time are considered jointly among the evaluated methods and budgets. Figures 3 and 5 in the revised manuscript explicitly present these trade-offs. Although individual baselines achieve better values on particular metrics, none simultaneously improves upon PAFM’s reported configurations across all three objectives.

Within the low-latency methods, PAFM offers a particularly strong physics–distribution balance. On Dynamic Stall, it achieves lower residual and WD than FM-OT and PBFM at every shared sampling budget. On Darcy Flow, it achieves the lowest residual among the evaluated millisecond-scale generators while substantially improving both pressure and permeability WD over PBFM. Methods that attain still lower residuals incur much larger inference costs. PAFM therefore extends the range of physically consistent, distributionally accurate generation achievable with inexpensive network inference.

W2: Novelty beyond learning physical constraints.

PAFM transforms physical consistency from a constraint imposed only at the final output into a guiding principle for every generation stage, enabling a small number of physically aligned transitions to achieve both distributional fidelity and physical consistency. Our key insight is to reorganize the generative process so that physical consistency is incorporated into every finite generation stage, instead of being enforced only at the final generation endpoint.

In conventional physics-constrained generative modeling, distributional evolution and physical consistency are often addressed separately: the generative model progressively transports samples toward the data distribution, while physical constraints are imposed through additional training objectives or enforced on the final generated samples. For multi-step flow-based generation, this creates a fundamental challenge: although the final endpoint can be physically constrained, the intermediate stages are not necessarily guided by physically compatible target states. Incorporating physical constraints throughout the generation process can therefore require additional physical evaluations and supervision, while balancing distributional fidelity and physical consistency.

PAFM takes a different perspective. We decompose each finite generation stage into two compatible operations: first, advancing the sample toward the data distribution in free physical coordinates; and second, completing the resulting state through known physical relations. Consequently, the physically completed next state serves not only as the endpoint of the current transition but also as the starting point of the subsequent transition. This constructs a sequence of physically compatible teacher endpoints, allowing physical consistency to guide the generative progression stage by stage, rather than being addressed only after the entire generation process.

This reformulation also provides a practical trade-off between generation quality, physical consistency, and computational cost. Instead of relying on a long sequence of unconstrained generative updates followed by physical correction, PAFM uses a small number of finite transitions, each guided by a physically completed target. We further distill the physical completion process into a lightweight transition recovery network, reducing the need for physical solves between learned transitions.

We believe this stage-wise physical alignment offers a broader research direction for physics-constrained generative modeling: rather than treating physical consistency as an auxiliary objective or a final-stage correction, it can be used to construct the targets and organize the progression of the generative process itself. The results across three benchmarks demonstrate the effectiveness of this formulation in achieving a favorable balance between distributional fidelity, physical consistency, and inference efficiency.

AuthorsOpenReview ↗

[2/3] Response to Reviewer hqpE

Q1: Why generate complete fields rather than generate only free variables and apply terminal completion? Do milliseconds matter?

We agree that generating only the free variables followed by terminal completion (TC) is a simple and competitive choice when an accurate, inexpensive completion operator is available at deployment. We have expanded this comparison in the revised manuscript by adding Dynamic Stall to Appendix F.1, Table 14, together with an analysis of the advantages and deployment requirements of both approaches.

However, PAFM addresses a more challenging task: directly generating complete, coupled physical fields while embedding physical consistency into the construction of its training targets. Our key insight is to make physics a guiding principle for every finite generation stage, rather than a constraint imposed only at the final output, enabling a small number of physically aligned transitions to achieve both distributional fidelity and physical consistency.

The new Dynamic Stall comparison directly examines the choice of generated variables. TC generates four free fields and completes pressure and wall shear at the end; PAFM generates the complete six-field state. Despite generating more channels, PAFM achieves 20.3% lower WD and 28.0% lower JS (Table 14). Reducing the generated state dimension therefore does not automatically improve distributional accuracy.

To isolate the effect of intermediate recovery, we also added a controlled comparison using the same frozen complete-state BASE (§5.2.4; Appendix F.2, Table 15). Here, TC denotes terminal recovery on that shared generator:

Shared BASE configurationTotal REWD × 10⁴JS × 10⁴
BASE + exact TC8.3460 × 10⁻¹⁶5.65274.1914
PAFM + exact TC8.3537 × 10⁻¹⁶5.59633.8993
BASE + learned TC7.9336 × 10⁻⁹5.65464.1916
PAFM7.4617 × 10⁻⁹5.59433.8970

Identical exact terminal completion brings both residuals to essentially the same level, yet PAFM retains 1.0% lower WD and 7.0% lower JS. The learned terminal head also uses the same lightweight, 2,498-parameter architecture as TRD. These comparisons support the benefit of feeding recovered physical coupling into subsequent transitions, beyond obtaining a physically consistent final output.

Crucially, PAFM also separates training-time physical computation from deployment requirements. With the physics-free initialization evaluated in Appendix E, complete coupled fields can be generated using model weights, noise, conditions, and fixed preprocessing, without accessing or executing the physical completion operator. For Darcy TC, every new permeability realization requires a coefficient-dependent pressure solve; PAFM learns this coupling during training.

The existing Darcy comparison demonstrates why the runtime difference matters: 5.60 versus 19.60 ms/sample is a 3.5× speedup, even though the absolute saving is 14 ms. Such savings accumulate when generating large ensembles of fields. The revised discussion clarifies that TC can be preferable when completion is cheap and available, while PAFM is especially relevant when per-sample physical computation is costly or the required physical model and solver are unavailable at deployment.

AuthorsOpenReview ↗

[3/3] Response to Reviewer hqpE

Q2: Two-round interaction, stability, and RL/imitation learning.

We thank the reviewer for raising this concern. The two-round TRD training is specifically designed to account for the distribution shift caused by the interaction between BASE and learned recovery.

BASE is trained on physically completed teacher states, whereas deployment involves states produced by its own finite transitions. The first round (R0) therefore learns physical recovery from BASE-generated rollouts. However, once TRD is introduced, the recovered states become the inputs to subsequent BASE transitions, changing the distribution of states encountered during generation. The second round (R1) explicitly accounts for this feedback by refreshing recovery supervision using rollouts generated by BASE with R0 enabled. This brings the recovery training distribution closer to the states encountered during deployment. BASE remains frozen throughout, and only one recovery head is used at inference.

This design follows our central motivation: BASE learns the physically aligned transport, while TRD restores the physical coupling of states encountered along the actual learned rollout. Although freezing BASE and restricting recovery do not provide an unconditional stability guarantee, our experiments support their effective interaction. On Dynamic Stall, TRD reduces RE from 8.08×10−68.08\times10^{-6} to 7.46×10−97.46\times10^{-9}, while also improving WD (5.6425 to 5.5943) and JS (4.3254 to 3.8970). On Kolmogorov Flow, RE decreases by 84.6% with negligible changes in distributional metrics.

We will also consider reinforcement learning and imitation learning as potential alternatives for achieving adaptation within a single training stage, and add a discussion of these directions to a future version of the manuscript. We thank the reviewer for this constructive suggestion.

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

The paper proposes PAFM for physics-constrained generation. It builds teacher states by moving free variables along a heat bridge from noise to data and completing each state with a known physical map, then trains a deterministic network (BASE) to predict each next state and a small recovery network (TRD) to correct rollout drift. On Kolmogorov Flow, Dynamic Stall and Darcy Flow it reports much lower residuals than PBFM at matched step budgets with competitive distributional metrics.

Critical strengths

• Using physically completed states as regression targets is a simple way to make the transition target and the physics loss agree per sample, and it gives large residual reductions over PBFM at the same NFE (e.g., 97.8% on Dynamic Stall at N=20). • The appendix is thorough: solver-free initialization, teacher-path ablations, cost accounting, and a terminal-completion comparison.

Critical weaknesses

• The simplest alternative beats the method and is only in the appendix. Generating the free variables and applying D_c once at the end (Terminal Completion, Table 14) gives RE 1.95e-10 vs. 1.74e-2 on Kolmogorov and 0.0012 vs. 0.0399 on Darcy, with better WD in both, at +1.5 ms and +14 ms per sample. Dynamic Stall uses a pointwise completion (~0.1 ms) but TC is not reported. None of the benchmarks has an expensive D_c, so the solver-free motivation is never really tested. The Kolmogorov TRD is even described as a linear periodic head matching the Leray closure, i.e., a learned approximation of a projection that can be applied exactly. • BASE is a deterministic regression on stochastic interpolant transitions, so it learns the conditional mean and drops the conditional variance at each step (the authors' own Eqs. 17–20); flow-map methods avoid this by learning deterministic ODE maps. This shows up in the results: at N=1 the output is basically the conditional mean (Darcy WD_K 0.798, close to FM-OT's 0.837), and on Dynamic Stall at N=5 and 10 PAFM has worse SMSE and JS than both FM-OT and PBFM (SMSE 7.27 vs. 1.30 for PBFM at N=5). The main text mostly reports WD. It also means the claimed compatibility between transition and physics objectives holds per sample but not for the learned optimum, as the CAG analysis concedes. • Distributional fidelity is not better than an unconstrained flow. On Darcy at N=20, FM-OT has WD_p 0.00067 vs. 0.0025 and WD_K 0.076 vs. 0.131. All results come from a single training seed, with a separate model and per-task, per-budget grid selection for each N.

Questions for authors

• Why is Terminal Completion not a main baseline, and what does it give on Dynamic Stall? Is there a setting where D_c is expensive enough that avoiding it matters? • Why learn the Kolmogorov TRD instead of applying the exact Leray projection? • Can you quantify sample diversity across N (e.g., per-condition variance ratio or coverage)? • Were FM-OT and PBFM retrained with the same backbone and budget? How does DiffusionPDE do at its native step count rather than 1–20 steps? • Heat and Linear bridges give nearly identical WD and JS on Kolmogorov (0.1063 vs. 0.1088). What supports the heat bridge beyond the scale-aware metrics?

5 comments
AuthorsOpenReview ↗

[1/4] Response to Reviewer sbJM

Thank you for the careful assessment. We address W1–W3 in the first two comments and Q1–Q5 in the following two comments. All references use the revised manuscript's numbering.

W1: Terminal completion, computational advantage, and the scope of PAFM.

We have added Dynamic Stall to the original terminal-completion (TC) comparison in Appendix F.1, Table 14, together with a discussion of each benchmark. TC achieves lower residual and distributional errors on Kolmogorov. On Darcy, it achieves lower RE and permeability WD, while PAFM has lower permeability JS. The new Dynamic Stall results show distributional gains for PAFM, detailed under Q1.

We agree that, when an accurate and inexpensive physical completion operator is available at deployment, generating only the free variables followed by a single terminal completion is a simple and competitive approach. Table 14 confirms its advantages in physical consistency and distributional fidelity in some evaluated settings. In subsequent versions, we will give this comparison greater prominence in the main text and further clarify PAFM’s intended scope and its trade-offs with TC.

PAFM’s central contribution is a physics-aligned probabilistic generation framework that transforms physical consistency from a constraint on the final output into a guiding principle for the generative transport itself. TC applies physical completion after generation in the free-variable space. PAFM instead uses physical completion during training to construct complete, physically aligned teacher states, making distributional supervision and physical consistency compatible at each finite-transition target. TRD further distills physical recovery into a lightweight network that learns to restore the coupling of predicted states without undertaking an independent probabilistic transport task. The recovered complete state then feeds into subsequent transitions, allowing the restored physical coupling to influence further distributional evolution. Physical information therefore specifies both the constraints that final samples should satisfy and the evolution through which the generator approaches the target distribution. This is the central modeling distinction between PAFM and generation followed by terminal completion.

The framework also separates physical computation during training from the requirements of deployment, reducing the generator’s dependence on explicit physical knowledge and numerical tools. Physical completion supplies teacher transitions and recovery supervision, allowing the model to learn coupling relations that would otherwise require explicit computation. With an initialization that requires no physical completion, deployment needs only the model weights, random noise, necessary generation conditions, and fixed preprocessing. Complete coupled fields can then be generated through network forward passes, without obtaining, implementing, or executing the completion operators and numerical solvers used during training. For example, Darcy TC must solve for pressure for every new permeability realization using the governing equation, fixed source term, and boundary conditions. PAFM learns this coupling during training and can generate the coupled fields without repeating that solve at deployment.

PAFM consequently retains the inference benefit of incorporating physics during training: it avoids additional physical solver or projection calls during the learned rollout. The existing Darcy experiment already demonstrates a meaningful computational advantage. PAFM takes 5.60 ms/sample versus 19.60 ms/sample for TC: a 3.5× speedup and 71.4% less sampling time. Although the absolute difference is 14 ms per sample, practical deployment may require very large ensembles of fields, making the savings from avoiding a coefficient-dependent PDE solve for every realization substantial.

PAFM thus converts explicit physical-computation dependencies into a deployable learned generative capability while reducing per-sample solution costs. We recognize that TC may be the preferable practical choice when completion is inexpensive and available. For complex physical coupling, costly per-sample solves, or deployment environments without access to the full physical model and solver, PAFM offers a promising generative framework. The current experiments establish its feasibility. Given the computational resources required for broader evaluation, we will add further assessments of computational benefits on more complex physical systems to the appendix in subsequent versions.

AuthorsOpenReview ↗

[2/4] Response to Reviewer sbJM

W2: Conditional averaging, sampling budget, and target compatibility.

We do not claim that PAFM achieves faithful one-step generation. The N=1N=1 setting evaluates an extreme sampling budget and exposes the averaging effect discussed in our analysis. We explicitly identify conditional averaging under finite-transition regression and formalize its physical consequence as the constraint averaging gap (CAG): even when individual teacher targets satisfy the physical relations, their conditional mean may violate nonlinear constraints (Appendices A–B; Eq. (28)).

Our analysis also provides a principled mitigation through finer transitions. Appendix B.2 relates CAG to constraint curvature and conditional covariance, and establishes, under the stated local regularity assumptions,

CAG⁡t,h(s)≤εdec,c+O(h2),\operatorname{CAG}_{t,h}(s)\leq \varepsilon_{\mathrm{dec},c}+O(h^2),

where hh is the transition length and εdec,c\varepsilon_{\mathrm{dec},c} is the physical-completion error floor (Eq. (31)). The conditional variance averaged out by each finite transition also scales as O(h2)O(h^2). Increasing the sampling budget to learn shorter transitions therefore reduces both local variance averaging and its contribution to physical inconsistency.

The experiments support this analysis. On Dynamic Stall, increasing NN from 5 to 20 reduces JS from 8.9274×10−48.9274\times10^{-4} to 3.8970×10−43.8970\times10^{-4} and SMSE from 7.2737×10−57.2737\times10^{-5} to 1.4513×10−51.4513\times10^{-5}, improvements of 56.3% and 80.0%, respectively. At N=20N=20, PAFM achieves lower JS than both FM-OT and PBFM; PBFM retains better SMSE, as explicitly reported in the main text and Table 2.

Further evidence comes from the direct diversity measurements reported in Q3. As NN increases from 1 to 20, Dynamic Stall’s variance ratio rises from 0.0137 to 0.6716, and coverage increases from 34.04% to 83.92% (Appendix K.1, Table 27). These results demonstrate substantial improvement in conditional variability with finer transitions, consistent with the theoretical analysis.

Our compatibility statement concerns the shared, physically completed teacher targets. The CAG analysis explicitly characterizes what happens when deterministic regression averages those targets and motivates how to mitigate the resulting gap. These are complementary parts of the contribution: constructing compatible supervision, identifying the averaging mechanism, and providing theoretical and empirical support for its mitigation through finer transitions.

W3: Marginal fidelity, single-seed results, and budget-specific selection.

We agree that FM-OT, a physically unconstrained flow-matching baseline, achieves better marginal distributional metrics on Darcy Flow. PAFM aims to achieve physical consistency while preserving distributional fidelity, so the relevant comparison considers both aspects of the generated fields.

In Darcy Flow, pressure and permeability must satisfy a sample-wise PDE. Randomly re-pairing their fields would preserve both marginal distributions while generally violating this coupling. At N=20N=20, PAFM achieves substantially stronger physical consistency than FM-OT, with 98.94% lower RE, but higher marginal WD. This trade-off is not inevitable: compared with the physics-constrained PBFM, PAFM simultaneously improves RE, WDp\mathrm{WD}_p, and WDK\mathrm{WD}_K by 95.3%, 83.5%, and 75.3%, respectively (Table 3). On Dynamic Stall, PAFM also achieves better WD than FM-OT (Table 2).

To further assess the reproducibility of these results, we have added the following three-seed evaluation at N=20N=20 on Darcy Flow, our benchmark with globally coupled PDE constraints.

MetricPAFM† (mean ± SD)
RE ↓0.040296 ± 0.006766
WDp\mathrm{WD}_p ↓0.00220356 ± 0.00052078
WDK\mathrm{WD}_K ↓0.13301397 ± 0.00274996
WDlog⁡K\mathrm{WD}_{\log K} ↓0.08112743 ± 0.00145234
JSp\mathrm{JS}_p ↓0.00060968 ± 0.00012886
JSK\mathrm{JS}_K ↓0.00432972 ± 0.00018389
JSlog⁡K\mathrm{JS}_{\log K} ↓0.00451482 ± 0.00019162
Valid ↑1.0000 ± 0.0000
IT (ms/sample) ↓5.628014 ± 0.025520

These results support stable performance across seeds: physical residuals remain low, permeability distribution metrics and inference time vary little, and validity remains 100%. PAFM retains a substantial physical-consistency advantage over FM-OT and PBFM, supporting the robustness of the reported gains.

Because each dataset and sampling budget requires separate training, a complete multi-seed evaluation across all settings incurs substantial computational cost. We will extend these evaluations to the remaining datasets and sampling budgets in subsequent versions of the paper.

AuthorsOpenReview ↗

[3/4] Response to Reviewer sbJM

Q1: Terminal completion on Dynamic Stall and its practical relevance.

We have added Dynamic Stall to the original TC comparison (Appendix F.1, Table 14). The generators have different state dimensions: TC evolves the four free fields and completes pressure and wall shear at the end; PAFM's BASE takes the complete six-field state as input and predicts increments for all six fields. PAFM achieves 20.3% lower WD and 28.0% lower JS, with RE 7.5×10−97.5\times10^{-9} and comparable sampling time (6.25 versus 6.30 ms/sample), demonstrating the benefit of learning coupled complete-field transitions.

The separate recovery-placement study (§5.2.4; Appendix F.2, Table 15) fixes the same six-field BASE, N = 20, prior, normalization, and test set. Here, TC denotes terminal recovery on that complete-field generator:

Shared BASE experimentTotal REWD × 10⁴JS × 10⁴
BASE + exact TC8.3460 × 10⁻¹⁶5.65274.1914
PAFM + exact TC8.3537 × 10⁻¹⁶5.59633.8993
BASE + learned TC7.9336 × 10⁻⁹5.65464.1916
PAFM7.4617 × 10⁻⁹5.59433.8970

Identical exact TC equalizes terminal consistency, yet PAFM retains 1.0% lower WD and 7.0% lower JS, supporting the effect of intermediate recovery on subsequent transport (Eq. (14)). Against learned TC, WD/JS improve by 1.1%/7.0% for approximately 0.45 ms/sample extra. The computational advantages of PAFM are discussed under W1.

Q2: Why learn Kolmogorov TRD rather than apply exact Leray projection?

We agree that exact Leray projection is the better choice for Kolmogorov Flow when the physical operator is available during generation. We include this benchmark specifically to demonstrate that projection-based physical corrections can also be distilled into PAFM’s recovery network. TRD substantially reduces divergence errors while largely preserving distributional fidelity, showing that the same framework accommodates projection-based constraints alongside nonlinear constitutive relations and global PDE coupling. This supports PAFM’s generality across different forms of physical constraints. We have clarified this motivation in Appendix F.1.

Q3: Sample diversity across sampling budgets.

New Appendix K.1, Table 27, and §5.2.4 report variance ratio (VR, ideal 1) and pointwise 5–95% interval coverage (nominal 90%) across all three benchmarks. The complete results are reproduced below; each entry reports VR / coverage.

Task / FieldNNPAFMPBFMFM-OT
Dynamic Stall10.0137 / 34.04%0.4002 / 75.12%0.2603 / 73.93%
50.2155 / 67.26%0.7255 / 79.04%0.5557 / 82.94%
100.5079 / 79.68%0.8515 / 82.53%0.6445 / 85.74%
200.6716 / 83.92%0.9253 / 84.61%0.6984 / 87.41%
Kolmogorov200.8852 / 86.42%0.8148 / 84.38%0.8469 / 85.52%
Darcy pp16.272×10−56.272\times10^{-5} / 1.42%5.6287 / 92.24%†^{\dagger}0.0028 / 9.55%
50.1606 / 54.45%0.6225 / 39.36%0.5042 / 78.36%
100.3940 / 73.38%0.8243 / 56.73%0.6950 / 84.31%
200.5884 / 82.43%0.9565 / 74.91%0.8166 / 86.93%
Darcy KK14.150×10−64.150\times10^{-6} / 0.43%1.5515 / 0.00%†^{\dagger}0.0004 / 3.11%
50.1636 / 64.35%0.0105 / 19.16%0.4185 / 77.42%
100.3852 / 78.74%0.0832 / 60.04%0.6148 / 84.00%
200.6054 / 84.74%0.2537 / 77.88%0.7595 / 86.95%
Darcy log⁡K\log K11.104×10−51.104\times10^{-5} / 0.43%—0.0005 / 3.11%
50.3199 / 64.35%——
100.5798 / 78.74%——
200.7630 / 84.74%——

†^{\dagger} At N=1N=1, every PBFM Darcy sample contains nonpositive permeability. Dashes indicate invalid log⁡K\log K statistics due to nonpositive permeability values.

Finer transitions improve PAFM's variability, although PBFM retains better Dynamic Stall VR. At N = 20, Kolmogorov PAFM obtains 0.8852/86.42%, versus PBFM's 0.8148/84.38% and FM-OT's 0.8469/85.52%. For Darcy K, PAFM improves over PBFM (0.6054/84.74% versus 0.2537/77.88%), while FM-OT remains better (0.7595/86.95%).

Comments on the paper

1 comment

Dear reviewers,

This is a friendly reminder to read the authors’ rebuttals and post your responses by Nov 18. Please briefly indicate which concerns have been addressed and which remain unresolved. Even if your overall assessment remains unchanged, please provide a brief response to the authors. Please also update your reviews and scores as appropriate. Thank you for your time and effort!

AC.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.