acceptodds
Withdrawn from ICLR 2027

PMosFM: Preconditioned Manifold Matching for One-Step Physics-Constrained Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Physics-constrained flow matching seeks to preserve a target distribution while respecting nonlinear physical constraints. Extending physics-constrained generation from multi-step sampling to a one-step finite-interval map requires learning the full constrained transport in a single regression problem, which can become poorly conditioned due to residual geometry and anisotropic interpolation statistics. To address this issue, we introduce Preconditioned Manifold one-step Flow Matching (PMosFM), a preconditioned manifold matching framework for one-step physics-constrained generation. An exact residual chart parameterizes the feasible degrees of freedom, removing the corresponding residual-loss branch and terminal residual unrolling. A geometric preconditioner rescales feasible coordinates according to the geometry induced by physical decoding, while an interpolation-state preconditioner reduces anisotropy in the states presented to the network. Then a two-time objective combines velocity supervision with consistency between decoded physical endpoints. Our framework characterizes residual-normal curvature, separates geometric and statistical factors in local regression conditioning, and relates physical flow-map defects to endpoint distribution error. Matched ablations show that the precondition–manifold–matching construction reaches a fixed physical–distributional validation criterion with fewer optimizer updates and lower wall-clock training time. At inference, PMosFM replaces multi-step integration of the learned flow with one-step generation, requiring a single network evaluation followed by physical decoding. Experiments across representative PDE benchmarks show lower end-to-end sampling latency than the evaluated multi-step baselines at comparable physical and distributional fidelity.

Primary area:
generative models
Keywords:
flow matching, physics-constrained generation, one-step generative modeling, scientific machine learning, partial differential equations

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_5nFLOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents a one-step generator which meets physical constraints. A decoder is leveraged to map the generated results to physical world. The authors first analyze why ambient space residual penalties can cause ill-conditioning. The decoder is leveraged to remove the constraints and the conditioning issue is discussed.

The authors conduct experiments on some standard tests like Darcy, Burgers, Kolmogorov flow and etc, demonstrating that their framework can achieve lower error and inference cost.

Critical strengths

The paper combines one-step generator with physics constraint decoder, reducing the inference cost.

The experiments cover a wide of PDE and standard tests.

Critical weaknesses

The proposed method has limited novelty. Tackling constrained problem via coordinate mapping and perform generation in a latent space is a standard approach. The physical residual error metric in the experiment is kind of circular residual metric, since the proposed method is guaranteed to have zero error by construction.

The comparison misses some important one-step baselines, like consistency model/meanflow trained with the same network architecture and inferred with the same decoder.

Questions for authors

In Table 14, why does the RMS error rise a lot when only gamma changes? Were the decoders active in all rows? Gamma influences the generative model training, but not the decoder. Since the physical constraint is guaranteed by the decoder, why the gamma value other than 0.5 gives big RMS errors?

Reviewer_owKdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of physical validity and computational efficiency in generating plausible solutions for PDE-based physical fields by proposing a method with less expensive training, one-shot sampling and following the physical constraints. The training uses flow matching with consistency loss on decoded fields, along with preconditioning. Several PDE benchmarks are used for evaluation with constraint adherence, distribution distances, and efficiency-based metrics.

Critical strengths

The proposed method works with one-shot sampling and has physical constraints built into the representations. It reports increases in sampling efficiency compared to evaluated multi-step baselines while maintaining the quality.

Critical weaknesses

1)The methodological contribution beyond the existing work is not clearly established. Methods like SoFlow [1] already establish finite interval maps and a consistency objective. The contributions of the main additions, such as endpoint matching in physical space and preconditioning, to the overall improvement are not clearly established.

  1. A core objective of the paper is adhering to the physics constraints, but it assumes that the decoder is provided, which guarantees the physics constraints. This makes it unclear if the proposed gains come from the proposed learning method or the decoder. It is also unclear how the framework can be extended to problems where such a decoder is unavailable.

[1] Luo, Tianze, Haotian Yuan, and Zhuang Liu. "SoFlow: Solution flow models for one-step generative modeling." International Conference on Learning Representations. Vol. 2026. 2026.

Questions for authors

One-step models can lose diversity.

  • Is there any empirical analysis of the generation diversity?
  • Is it prone to memorization?
8 comments
AuthorsOpenReview ↗

Part 1: Responses to the Critical Weaknesses 1

We thank the reviewer for recognizing the efficiency of one-step generation and for raising important questions about the learning mechanism and the role of the physical decoder.

W1. The methodological contribution beyond existing work is not clearly established; SoFlow already provides finite-interval maps and a consistency objective.

SoFlow already provides finite-interval flow maps and velocity-anchored solution consistency. However, SoFlow evaluates consistency of the generative velocity ODE solution map, rather than the governing PDE residual Rh(x,c)=0R_{h}(x,c)=0. PCFM[1] applies multiple physics corrections during sampling, increasing inference cost. PBFM[2] optimizes PDE residuals during training and uses unrolling for terminal-residual evaluation, increasing training cost. PMosFM generates samples in one step, with Rh(Ψc(r),c)=0R_{h}(Ψ_{c}(r),c)=0 enforced by the decoder on the valid decoding domain. For the encoded constraints, PMosFM needs no additional sampling-time physics corrections, additional residual penalties, or terminal-residual unrolling during training.

With the same exact decoder, SoFlow trained in decoder coordinates has the same feasibility guarantee as PMosFM. PMosFM instead measures endpoint-matching errors in physical space and preconditions the decoder-induced metric and interpolation-state covariance. For endpoint coordinates aa and bb, coordinate-space consistency penalizes ‖a−b‖22‖ a-b‖_{2}^{2}, whereas PMosFM penalizes ‖Ψc(a)−Ψc(b)‖Mh2‖Ψ_{c}(a)-Ψ_{c}(b)‖_{M_{h}}^{2} in Eq. 12:

= (a-b)^{T}G̃_{c}(b)(a-b) + O(‖ a-b‖^{3}), G̃_{c}(b)=J_{Ψ_{c}}(b)^{T} M_{h}J_{Ψ_{c}}(b).$$ Coordinate errors with the same Euclidean norm can incur different physical-space penalties and produce different training gradients. $C_{c}$ is used to improve the conditioning of the decoder-induced metric. $P_{s,c}$ preconditions the covariance of interpolated states in Eq. 15. Neither $C_{c}$ nor $P_{s,c}$ is a scalar loss weight. Fig. 5 reports conditioning diagnostics. For an affine-isometric decoder, the two endpoint distances coincide, as expected. With the same decoder and feasible representation, PMosFM-native has lower energy distance than PMosFM–SoFlow across the reported benchmarks (Table 4; Darcy: 0.0043 vs. 0.0317). In the following PDE benchmarks, PMosFM and SoFlow both use one network evaluation, but PMosFM achieves lower RE, WD, and JS. | Benchmark | Metric | PIDM | ECI | PCFM | PBFM | PMosFM | SoFlow | | --- | --- | --- | --- | --- | --- | --- | --- | | Dynamic Stall | RE ×10⁶ ↓ | 2.96×10³ | 40.7 | 0.143 | 0.339 | **0.0421** | 468.7 | | Dynamic Stall | WD ×10⁴ ↓ | 179 | 11.1 | 4.01 | 1.81 | **0.842** | 712.1 | | Dynamic Stall | JS ×10² ↓ | 4.14 | 0.0570 | 7.46 | 0.680 | **0.0368** | 1.323 | | Dynamic Stall | NFE ↓ | 100 | 200 | 20 | 20 | **1** | **1** | | Dynamic Stall | IT [s] ↓ | 0.0866 | 0.432 | 3.91 | 0.0605 | 0.00338 | **0.00329** | | Darcy Flow | RE ↓ | **0.0220** | 3.05 | 4.19×10³ | 0.838 | 0.120 | 230.2 | | Darcy Flow | WD ×10² ↓ | 3.10 | 2.89 | 3.74 | 0.138 | **0.0490** | 7.297 | | Darcy Flow | JS ×10¹ ↓ | 3.18 | 2.82 | 0.199 | 0.256 | **0.0139** | 0.1284 | | Darcy Flow | NFE ↓ | 100 | 20 | 20 | 20 | **1** | **1** | | Darcy Flow | IT [s] ↓ | 2.05 | 0.122 | 1.33 | 0.101 | 0.0178 | **0.0174** | | Burgers | RE ×10¹ ↓ | 1.57×10³ | 0.818 | 0.473 | 0.307 | **0.0435** | 32.35 | | Burgers | WD ×10² ↓ | 248 | 2.55 | 2.73 | 4.02 | **0.129** | 7.481 | | Burgers | JS ×10² ↓ | 31.5 | 0.809 | 0.884 | 1.43 | **0.0320** | 11.17 | | Burgers | NFE ↓ | 20 | 20 | 20 | 20 | **1** | **1** | | Burgers | IT [s] ↓ | 0.0354 | 0.0902 | 0.134 | 0.0913 | 0.00237 | **0.00236** | *Note: RE denotes relative error, WD Wasserstein distance, and JS Jensen–Shannon divergence; lower is better.* [1] U. Utkarsh et al. “Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints.” *Advances in Neural Information Processing Systems 38, 2025.* [2] G. Baldan et al. “Physics vs Distributions: Pareto Optimal Flow Matching with Physics Constraints.” *International Conference on Learning Representations ,2026.*
AuthorsOpenReview ↗

Part 1: Responses to the Critical Weaknesses 2

W2. The decoder already guarantees the constraints. Do the gains come from the learning method or the decoder, and what happens when such a decoder is unavailable?

With a fixed decoder, PMosFM improves distributional accuracy. We compare an unlearned prior, a Gaussian fitted to the training coordinates, and PMosFM using the same task-specific decoder. Within the benchmarks, the three samplers attain comparable encoded-constraint residuals but different distributional accuracy. Relative to the fitted-Gaussian baseline, PMosFM reduces energy distance (ED) by 53.8%, 70.9%, and 92.6%. The shared-decoder results show that learned transport improves agreement with the target distribution while maintaining comparable encoded-constraint residuals.

BenchmarkMethod (same ΨcΨ_{c})Encoded residual RMS ↓ED ↓WD (×10⁻²) ↓Latency (ms) ↓
Dynamic Stall (algebraic EOS)Prior-only4.62×10⁻¹⁷0.01810.02710.012
Dynamic Stall (algebraic EOS)Training-fitted Gaussian2.33×10⁻¹⁷0.00780.01810.074
Dynamic Stall (algebraic EOS)PMosFM (learned)2.35×10⁻¹⁷0.00360.00840.519
Burgers (mass conservation)Prior-only<10⁻¹³0.31525.6200.015
Burgers (mass conservation)Training-fitted Gaussian<10⁻¹³0.12851.4500.068
Burgers (mass conservation)PMosFM (learned)<10⁻¹³0.03740.1290.482
Darcy Flow (implicit solver)Prior-only1.12×10⁻³0.24728.1655.12
Darcy Flow (implicit solver)Training-fitted Gaussian0.98×10⁻³0.05841.8425.19
Darcy Flow (implicit solver)PMosFM (learned)0.955×10⁻³0.00430.04905.64

PMosFM uses decoded physical endpoints in the training loss. With the decoder and feasible representation fixed, PMosFM-native's lower ED relative to PMosFM–SoFlow indicates a distributional benefit from the learning design (Table 4). The decoder-induced metric makes the training gradient sensitive to physical-space errors, while preconditioning acts on the metric and interpolation-state covariance (Eq. 12; Fig. 5).

An implicit decoder trades numerical accuracy for decoding cost. An implicit-decoder study below documents the numerical accuracy–cost trade-off: increasing KK from 16 to 256 reduces the independent relative residual error from 0.0385 to 0.00012, while per-sample decoding latency rises from 0.62 to 5.12 ms. Finite-iteration decoding provides approximate constraint satisfaction, not exact feasibility, with residual accuracy depending on the solver budget. The independent relative residuals and the encoded RMS values in the shared-decoder table are different metrics. End-to-end sampling cost must include numerical decoding. The PMosFM formulation requires a suitable feasible representation and a usable explicit or implicit decoder.

Budget KKRelative residual RE ↓Relative pressure error ↓Decode latency ↓
160.0385±0.00350.0124±0.00120.62 ms
320.0084±0.00090.0031±0.00040.98 ms
640.0019±0.00020.0008±0.00011.65 ms
1280.00048±0.000050.0002±0.000032.84 ms
2560.00012±0.000020.00004±0.000015.12 ms
AuthorsOpenReview ↗

Part 2: Response to the Question

Part 2: Responses to the Questions on Diversity and Memorization

We thank the reviewer for asking whether one-step generation preserves sample diversity and whether PMosFM is prone to memorization.

Q1. One-step models can lose diversity. Is there any empirical analysis of generation diversity?

We share the reviewer's concern about diversity loss and assess samples from models with fixed weights. Variance and covariance spectra measure variation, while calibrated recall measures reference coverage.

Darcy Flow retains 80.2% of the reference physical variance and matches the validation-reference recall at 94.5%. The lower auxiliary participation rank indicates fewer effective directions of variation.

For Burgers, reference trajectories retain 2.27% of raw variance after encoding and decoding with the 12-mode Cole–Hopf chart. PMosFM reaches 96.8% of the encode–decode reference recall and 95.3% reference proximity. Generated samples retain 6.7% of raw variance.

Dynamic Stall retains 84.3% of the reference physical variance. Raw reference samples mix geometries, whereas generation uses a fixed decoder condition. Recall against the condition-matched reference is 93.8%.

BenchmarkVG/VRV_{G}/V_{R}(rG,rR)(r_{G},r_{R})Spectrum L1L_{1}Support Recall
Dynamic Stall0.843(9.24,1.82)1.15693.8
Darcy Flow0.802(8.92,11.89)0.17794.5
Burgers0.067(1.82,2.66)0.64596.8

Note: G/R: generated/reference; VV: physical covariance trace; rr: participation rank; L1L_{1}: spectral distance.

Q2. Is the method prone to memorization?

The memorization tests show generated near-copy counts at or below held-out counts. We flag a sample when the distance to the nearest training sample falls below an independently calibrated 5th-percentile threshold. Held-out samples provide the baseline; 128/128 exact replays are flagged, confirming sensitivity to exact training copies.

For Burgers, encoding and decoding can change a copied trajectory enough to exceed the physical-space threshold. Physical screening detects 4/128 perturbed replays, whereas coordinate screening detects 128/128 of the 1%-SD perturbed controls and flags 0/128 generated samples. The large coordinate distance ratio q=30.852q=30.852 indicates distribution shift; generalization should be assessed using distributional accuracy on held-out data.

Dynamic Stall uses non-overlapping training and test conditions, so we calibrate the copying screen with condition-matched controls. Physical and feature screening detect 128/128 exact and perturbed replays and flag 0/128 generated samples. Coordinate distances across conditions are omitted because the decoder charts depend on the condition. The byte-level audit found no identical targets.

The near-copy detection results apply only to the tested checkpoints and calibrated replay controls.

BenchmarkSpaceqqG flagsR flagsReplayPerturbed
Dynamic StallPhysical1.6780/1286/128128/128128/128
Dynamic StallFeatures1.5890/1286/128128/128128/128
Darcy FlowPhysical0.9592/12811/128128/128128/128
Darcy FlowFeatures0.9911/12810/128128/128126/128
Darcy FlowCoordinates0.9827/1288/128128/128128/128
BurgersCoordinates30.8520/1288/128128/128128/128

Note: qq: generated/held-out median nearest-neighbor distance ratio; G/R flags: near-copy counts.

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

This paper focus on the physics-constrained generative models to generate physical fields that match a target distribution and satisfy prescribed constraints. The authors propose a preconditioned manifold one-step flow matching framework for one-step physics-constrained generation. For the objective function, it is a combination of an ordinary flow matching and manifold matching.

Critical strengths
  1. One-step physics-constrained generation.
  2. Preconditioned-manifold-matching strategy.
  3. Feasibility, conditioning and transport.
  4. The numerical results look good.
Critical weaknesses
  1. The writting is poor, which is not easy to follow until Line 225.
  2. The preconditioning is very similar as data preprocessing.
  3. Some notations are confusing.
Questions for authors
  1. Figure 1 is confusing. There are some notations used before definition.
  2. When reading the Subsections 3.1 and 3.3, I have no ideas what you want to do? Would you please explain more about the connections between these parts?
  3. I am not sure the effect of the introduction of the preconditioning. Ps,c(r−ms,c)P_{s,c}(r-m_{s,c}) is a linear expression of rr, is it possible to contribute to a part of nn? In addition, Ps,c(r−ms,c)P_{s,c}(r-m_{s,c}) is very similar as the data preprocessing. Set r′=Ps,c(r−ms,c)r'=P_{s,c}(r-m_{s,c}) and use r′r' as a new input of the flow matching. Would you please have a try?
  4. It is not good to use mm as the dimesion, where you use ms,cm_{s,c} as the mean.
  5. There is no definitio of sg(⋅)sg(\cdot).
4 comments
AuthorsOpenReview ↗

Part 1: Responses to the Critical Weaknesses 1 and 3

We thank the reviewer for the constructive feedback. We have revised the presentation to clarify the motivation, methodological connections, and notation.

W1. The writing is poor, which is not easy to follow until Line 225.

We agree that PMosFM's motivation and main idea needed clearer explanation. The revised opening sections present the motivation and generation procedure before the mathematical details.

The Introduction now states two goals: physical feasibility and distributional fidelity. We explain the computational costs of sampling-time corrections in PCFM[1] and residual-based training in PBFM[2]. PMosFM is introduced through a sampling procedure: draw a source coordinate, apply one learned transport update, and decode the result into a physical field. We also clarify why feasible-coordinate transport benefits from preconditioning and how finite-interval endpoint consistency complements velocity supervision.

The Methodology now follows a clearer progression. Section 3.1 defines the feasible set and target data distribution, Section 3.2 explains the feasibility gap in ambient-space flow matching, and Section 3.3 constructs the PMosFM generator. Figure 1 illustrates the motivation, while Fig. 2 shows the training and sampling procedures.

These revisions clarify the motivation, main algorithm, and connections among the proposed components while preserving the mathematical formulation and results.

W3. Some notations are confusing.

We agree and have revised the notation throughout. The interpolation mean is now rˉs,c\bar r_{s,c} (previously ms,cm_{s,c}), and the intrinsic dimension is now denoted by dMd_{\mathcal M}, distinct from the physical state dimension dd, kept distinct from the chart center μc\mu_c and from the sampling time t∈[0,1]t\in[0,1]. The terms “student” and “chord” have been removed, chart-related vocabulary has been reduced in favor of feasible coordinates and parameterizations, and sg(⋅)\mathrm{sg}(\cdot) is defined at its first appearance. Figure 1 now retains only seven mathematical symbols, all defined before use.

These changes preserve the core equations while making the notation easier to follow.

AuthorsOpenReview ↗

Part 1: Responses to the Critical Weaknesses 3

W2. The preconditioning is very similar as data preprocessing.

We thank the reviewer for this observation. PMosFM preconditions the covariance spectrum seen by the network during training. As analyzed in Preconditioned Flow Matching (PFM) [3], anisotropic interpolation covariance can slow learning along low-variance directions while high-variance directions fit more quickly.

PMosFM addresses this imbalance using

Ps,c=(Σs,c+εPI)−1/2,ζs,c=Ps,c(rs−rˉs,c),P_{s,c}=(\Sigma_{s,c}+\varepsilon_P I)^{-1/2}, \qquad \zeta_{s,c}=P_{s,c}(r_s-\bar r_{s,c}),

where Σs,c\Sigma_{s,c} is the interpolation-state covariance. We estimate and freeze the statistics before transport training. Each sampled time ss and condition cc selects the corresponding input matrix. The transformed covariance is

Cov⁡(ζs,c∣s,c)=Ps,cΣs,cPs,c⊤.\operatorname{Cov}(\zeta_{s,c}\mid s,c) =P_{s,c}\Sigma_{s,c}P_{s,c}^{\top}.

Following matrix preconditioning in scientific computing, this training-time mapping reduces covariance spectral anisotropy to improve regression conditioning. The original interpolation states and physical objective remain unchanged. In the revised manuscript, Fig. 6 links improved conditioning to faster convergence and lower low-variance mode error.

An ablation study evaluates preconditioning on an anisotropic Gaussian task with an identity decoder. Static z-score and the learned diagonal affine layer share one map across interpolation times; PMosFM uses the map indexed by ss.

Input mappingDependence on interpolation timeκ at s=0.9Weak-direction error ↓Strong-direction error ↓
IdentityNone81.41±0.460.1028±0.01250.1872±0.0087
Static target z-scoreFixed across all s133.16±0.740.1528±0.00720.2742±0.0138
Learned diagonal affineFixed across all s74.79±7.120.0972±0.01610.1789±0.0100
PMosFM, εP=0.05ε_{P}=0.05Indexed by current s5.63±0.030.0716±0.00490.1785±0.0127

PMosFM reduces the covariance condition number from 81.4181.41 to 5.635.63, compared with 74.7974.79 for the learned diagonal affine layer. The calibrated map achieves lower weak-direction relative endpoint RMSE with strong-direction accuracy comparable to the learned affine baseline. These gains use zero additional trainable affine parameters and support explicit covariance preconditioning for more balanced directional learning.

Figure 17 in the revised manuscript's Additional Ablations shows network-input covariance conditioning across interpolation times using εP=10−5\varepsilon_P=10^{-5}; the table above uses 0.050.05.

Figure 18 in the revised manuscript compares weak- and strong-direction RMSE and weak-direction relative variance error. Time-dependent maps reduce weak-direction error, with aggressive whitening increasing strong-direction error.

Geometric preconditioning CcC_c addresses decoder-induced physical sensitivity. The local theoretical bound for physical-endpoint regression is

κ(HAGN)≤βακ(Ps,cΣs,cPs,c⊤).\kappa(H_A^{\mathrm{GN}})\leq\frac{\beta}{\alpha} \kappa(P_{s,c}\Sigma_{s,c}P_{s,c}^{\top}).

The metric factor β/α\beta/\alpha and covariance factor identify the complementary roles of CcC_c and Ps,cP_{s,c}. Together, geometric and statistical preconditioning improve the optimization geometry of feasible-coordinate transport.

[1] U. Utkarsh et al. “Physics-Constrained Flow Matching: Sampling Generative Models with Hard Constraints,” Advances in Neural Information Processing Systems 38 (NeurIPS 2025).

[2] G. Baldan et al. “Physics vs Distributions: Pareto Optimal Flow Matching with Physics Constraints,” International Conference on Learning Representations (ICLR 2026).

[3] S. Ahamed et al. “Preconditioned Flow Matching,” arXiv preprint arXiv:2603.02337 (2026).

AuthorsOpenReview ↗

Part 2: Response to the Question 1, 2, 4, 5

We thank the reviewer for the helpful questions and address them individually below.

Q1. Figure 1 is confusing. There are some notations used before definition.

We agree and have revised Fig. 1 to clarify how the three methods handle physical constraints. PCFM applies repeated physics corrections during sampling, while PBFM incorporates physical residuals into training through ConFIG and unrolling. PMosFM conditions interpolation-state inputs, matches decoded endpoints during training, and performs one-step transport followed by physical decoding. We removed unnecessary mathematical symbols and clarified the remaining notation in the manuscript text and caption.

Figure 2 separately illustrates the PMosFM framework, while Section 3.3 explains the training and sampling procedures.

Q2. When reading Subsections 3.1 and 3.3, I have no idea what you want to do. Would you please explain more about the connections between these parts?

We thank the reviewer for noting the need to clarify how Sections 3.1–3.3 connect feasibility to generation, and have revised the Methodology accordingly.

Section 3.1 specifies the admissible physical states through the constraint set

Zc=x:Rh(x,c)=0.\mathcal Z_c=\\{x:R_h(x,c)=0\\}.

The target data distribution assigns probability over this feasible set. Section 3.2 explains why ambient-space flow matching does not guarantee physical feasibility.

Section 3.3 constructs the PMosFM generator. A feasible decoder Ψc\Psi_c maps valid intrinsic coordinates to physical fields satisfying

Rh(Ψc(r),c)=0R_h(\Psi_c(r),c)=0

for coordinates in the decoder's valid domain. During training, observed fields are encoded as target coordinates r1r_1 and paired with source coordinates r0r_0 for velocity supervision. Preconditioning improves transport conditioning, while physical-endpoint consistency trains finite-interval predictions.

At inference, PMosFM applies one transport update, r^1=Tθ0,1(r0,c)\widehat r_1=T_\theta^{0,1}(r_0,c), and decodes the endpoint as x^1=Ψc(r^1)\widehat x_1=\Psi_c(\widehat r_1). The decoder represents feasible physical states, while the learned transport approximates the target distribution over those states.

Q4. It is not good to use mm as the dimension, where you use ms,cm_{s,c} as the mean.

Thank you for pointing out the ambiguity. We now use dMd_{\mathcal M} for the intrinsic dimension and rˉs,c=E[rs∣c]\bar r_{s,c}=\mathbb E[r_s\mid c] for the interpolation mean. We have applied these changes consistently throughout the main equations, loss normalizations, algorithms, notation table, and appendix proofs.

Q5. There is no definition of sg(⋅)\mathrm{sg}(\cdot).

Thank you for pointing this out. We now define sg⁡(⋅)\operatorname{sg}(\cdot) as the stop-gradient operator: the forward pass returns the input unchanged, while backpropagation assigns zero gradient. In the physical-endpoint objective, θˉ=sg⁡(θ)\bar\theta=\operatorname{sg}(\theta) blocks gradients through the reference endpoint.

AuthorsOpenReview ↗

Part 2: Response to the Question 3

Q3.1. I am not sure the effect of the introduction of the preconditioning. Ps,c(r−ms,c)P_{s,c}(r-m_{s,c}) is a linear expression of rr; is it possible to contribute to a part of NN?

Thank you for raising this question. For fixed s,cs,c, the affine transformation can be absorbed into a compatible first network layer:

WPs,c(r−rˉs,c)+b=(WPs,c)r+(b−WPs,crˉs,c).W P_{s,c}(r-\bar r_{s,c})+b =(W P_{s,c})r+(b-WP_{s,c}\bar r_{s,c}).

An implementation check confirms this equivalence to numerical precision. Inspired by PFM [3] and matrix preconditioning in scientific computing, PMosFM applies:

Ps,c=(Σs,c+εPI)−1/2,ζs,c=Ps,c(rs−rˉs,c),P_{s,c}=(\Sigma_{s,c}+\varepsilon_P I)^{-1/2},\qquad \zeta_{s,c}=P_{s,c}(r_s-\bar r_{s,c}),

to each training input at the sampled interpolation time ss and condition cc. Thus, the covariance seen by the network is

Cov⁡(ζs,c∣s,c)=Ps,cΣs,cPs,c⊤.\operatorname{Cov}(\zeta_{s,c}\mid s,c)=P_{s,c}\Sigma_{s,c}P_{s,c}^{\top}.

The calibrated matrices are fixed; during training, each input uses Ps,cP_{s,c} indexed by the sampled s,cs,c, conditioning the covariance seen by the network while leaving the distribution of rsr_s unchanged.

We compare the input mappings under a matched training setup. The table reports low-variance relative endpoint RMSE; parameter counts exclude the shared transport network.

MappingTime dependenceExtra trainable params.Low-variance RMSE ↓\downarrow
UnpreconditionedNone00.1028±0.01250.1028\pm0.0125
Learnable diagonal affineStatic240.0972±0.01610.0972\pm0.0161
Learnable affineTime-aware2640.0697±0.00810.0697\pm0.0081
Calibrated Ps,cP_{s,c}Time-aware00.0716±0.00490.0716\pm0.0049
Calibrated Ps,cP_{s,c} + learned correctionTime-aware1320.0562±0.0035\mathbf{0.0562\pm0.0035}

The calibrated preconditioner achieves comparable low-variance RMSE to the time-aware learned affine layer (0.0716 vs. 0.0697), without additional trainable parameters. A separate learned-correction variant further reduces RMSE to 0.0562. These results show that explicit covariance-aware preconditioning achieves comparable endpoint accuracy without adding trainable parameters.

Q3.2. In addition, Ps,c(r−ms,c)P_{s,c}(r-m_{s,c}) is very similar as the data preprocessing. Set r′=Ps,c(r−ms,c)r'=P_{s,c}(r-m_{s,c}) and use r′r' as a new input of the flow matching. Would you please have a try?

Thank you for this suggestion. The original presentation left the transformed network input and the calibration schedule for that map unclear. PMosFM already uses the proposed input:

Tθs,t(r,c)=r+(t−s)uθ(Ps,c(r−rˉs,c)⏟r′,s,t,c).T_\theta^{s,t}(r,c) = r + (t-s)u_\theta\left(\underbrace{P_{s,c}(r-\bar r_{s,c})}_{r'}, s, t, c\right).

Here, r′r' is the input to the velocity network. The transport map adds the predicted displacement to the original state rr, so the interpolation path, velocity targets, and physical decoder remain in the original intrinsic coordinates.

We estimate and freeze the preconditioning statistics from the training split before transport optimization. During training, each sampled interpolation time ss and condition cc selects its corresponding matrix Ps,cP_{s,c}. The calibrated matrices remain fixed throughout optimization.

This ablation study compares the time-dependent map with identity, target-fitted static z-score, and time-independent learned affine inputs under the same training protocol. The table reports weak-direction relative endpoint RMSE.

Input mappingNetwork input r′r'Time-dependent?Weak-direction RMSE ↓\downarrow
Identityrsr_sNo0.1028±0.01250.1028 \pm 0.0125
Target-fitted static z-scorePstatic(rs−rˉ1)P^{\mathrm{static}}(r_s-\bar r_1)No0.1528±0.00720.1528 \pm 0.0072
Time-independent learned affineLrs+bLr_s+bNo0.0972±0.01610.0972 \pm 0.0161
PMosFM time-dependentPs(rs−rˉs)P_s(r_s-\bar r_s)Yes0.0716±0.0049\mathbf{0.0716 \pm 0.0049}

PMosFM achieves a weak-direction RMSE of 0.0716±0.00490.0716\pm0.0049, compared with 0.0972±0.01610.0972\pm0.0161 for the learned static affine map, 0.1528±0.00720.1528\pm0.0072 for target-fitted z-score, and 0.1028±0.01250.1028\pm0.0125 for identity inputs. The result supports time-dependent conditioning for the tested interpolation path.

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

Related papers

Discussion (0)

Sign in to comment.