acceptodds
Under review as a conference paper at ICLR 2027

Physics-Embedded Flow Matching: Governing the Local Vector Field with a Known Physiological Relation

Abstract

Physiological signal translation often suffers from limited paired data, motivating the use of governing relationships between source and target measurements as additional supervision. Existing physics-informed models constrain supervised predictions, while their flow matching counterparts constrain reconstructed states, generated endpoints, or generic properties of the transport field. We introduce a physics-informed flow matching framework that instead imposes the physiological relationship directly on the learned local velocity field through a physical constraint. We derive a closed-form velocity-space residual that softly enforces physical consistency at each flow time, and we present a unified error analysis characterizing how physics regularization and multi-step integration can reduce prediction error. Using a learned two-element Windkessel model, we evaluate translation from virtual aortic flow and from measured bioimpedance to pressure against supervised and flow matching baselines with and without physics constraints. Our Physics Embedded Flow Matching (PEFM) method achieves the lowest waveform RMSE on both datasets. On measured data, it achieves lower RMSE than the baseline closest in physical consistency and lower physics residual than the baseline closest in RMSE. These results support the use of governing relationships to jointly improve predictive accuracy and physical consistency.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Physics-Informed Machine Learning, Flow Matching, Physiological Signal Translation

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

This paper presents Physics-Embedded Flow Matching (PEFM), a framework for paired physiological signal translation that incorporates known physiological relationships into the learned local velocity field. The method derives a physics residual in velocity space and combines it with the standard flow matching objective. For affine interpolation paths and linear physical operators, the paper establishes a connection between local velocity residuals along the generated trajectory and the physical consistency of the final prediction. It also provides an idealized analysis of how physics regularization and error cancellation during multi-step integration can improve prediction accuracy. The framework is evaluated on two datasets for blood pressure waveform reconstruction: one using simulated aortic flow and the other using a flow proxy derived from measured bioimpedance. Both experiments employ a two-element Windkessel model with learned parameters. PEFM achieves the lowest reported waveform RMSE and MAE among the evaluated methods on both datasets, while achieving physical consistency close to WPINN on the measured dataset.

Critical strengths

A key strength of this paper is the formulation of a mathematical representation that translates known physiological relationships into local velocity field regularization terms, thereby eliminating the need for full ODE trajectory unrolling during training. This approach establishes a link between local velocity residuals and endpoint physical consistency, providing a theoretical basis for the proposed scheme. The authors present detailed derivations in both the main text and the appendix, lending the paper a well-structured quality.

Critical weaknesses
  1. Evidence supporting the main argument is limited. While the introduction describes the proposal of a flow-matching framework, both validation experiments focus exclusively on blood pressure reconstruction based on the Windkessel relationship. Consequently, the framework's applicability across different physiological mechanisms remains unverified. Although the results demonstrate its utility for blood pressure reconstruction, they are insufficient to establish broader applicability to other physiological signal transformation tasks.

  2. The scale of the dataset used for practical application is limited. The real-world dataset comprises only seven subjects. In each data split, three subjects are allocated for training, two for validation, and two for testing. Despite the reporting of results based on five random seeds, the small sample size limits the persuasiveness of conclusions regarding inter-subject robustness and generalization capability.

Questions for authors

Questions 2 and 3 are my primary concerns, and the authors' responses are crucial for my final assessment.

  1. As stated in the introduction, can the proposed framework be extended to other physiological signals (e.g., ECG)? What control relationships would this require, and how should the existing physical constraints be adjusted accordingly?

  2. Within the scope of the current experimental setup (specifically blood pressure reconstruction), could the authors utilize additional datasets to strengthen the evaluation and assess the framework's robustness and generalization capabilities across different subjects group (like different age, BMI, etc.) and measurement conditions?

  3. Within the same scope as Q2, could the authors provide an assessment for downstream tasks to demonstrate the reliability and practical application value of the reconstructed blood pressure signal?

4 comments

We thank the reviewer for their constructive assessment of our formulation and theoretical analysis. We address (W1/Q1) the limited validation across physiological mechanisms by introducing electrophysiological and respiratory experiments, (W2/Q2) the small sample size and generalization concerns through additional datasets with different subject groups and a large-scale subject-disjoint evaluation, and (Q3) the clinical relevance of reconstructed BP through downstream SBP/DBP estimation.

(W1/Q1) Extension to other physiological signals.

PEFM applies whenever a known governing relationship connects source and target signals. For linear relationships, the corresponding operator can be incorporated directly into the velocity-space residual without changing the FM formulation. We demonstrate this on two additional physiological mechanisms.

Electrophysiology (EDGAR [1]). We reconstructed heart-surface potentials from torso measurements using the linear forward model

ytorso(t)=Hxheart(t),y_ {\mathrm{torso}}(t)=Hx_ {\mathrm{heart}}(t),

where HH is the geometry-dependent transfer matrix. We evaluated two cardiac preparations containing 205 and 927 signal windows, respectively.

EDGAR results. Test RMSE (mV, mean ± std over three seeds) and grouped Wilcoxon pp-values against PEFM, pooled over seeds. Grouping is by beat (u020515) or recording run (u180809).

Methodu020515 RMSE ↓u020515 pu180809 RMSE ↓u180809 p
Supervised0.293 ± 0.044<0.0010.757 ± 0.0360.031
FM0.218 ± 0.027<0.0010.721 ± 0.0460.031
WPINN0.306 ± 0.051<0.0010.841 ± 0.0450.031
PIFM-GN0.200 ± 0.0210.1460.646 ± 0.1660.156
PBFM0.217 ± 0.027<0.0010.696 ± 0.1650.156
PEFM0.196 ± 0.025–0.608 ± 0.116–

PEFM achieves the lowest mean RMSE on both preparations, improving upon FM by 10.1% and 15.7%, respectively. Improvements over three baselines reach nominal significance in both preparations, although seed-level significance is established only for u020515.

Respiratory mechanics (CCVW-ICU [2]). We reconstructed transpulmonary pressure from airway flow using the single-compartment equation

PL(t)=RQ(t)+EV(t)+P0.P_ L(t)=RQ(t)+EV(t)+P_ 0.

We evaluated 276 breaths from seven mechanically ventilated ICU patients.

CCVW-ICU results. Test RMSE (cmH₂O, mean ± std over three seeds) and patient-grouped Wilcoxon pp-values against PEFM (learned), pooled over seeds.

MethodRMSE ↓p-value
Supervised3.517 ± 0.3630.062
FM3.437 ± 0.6610.062
WPINN (oracle)3.654 ± 0.4270.062
WPINN (learned)2.704 ± 0.8900.312
PIFM-GN3.532 ± 0.7800.062
PBFM (oracle)3.389 ± 0.8640.094
PBFM (learned)3.202 ± 1.0450.031
PEFM (oracle)2.600 ± 0.502–
PEFM (learned)2.382 ± 0.892–

PEFM achieves the lowest mean RMSE, improving upon FM by 30.7% and WPINN (learned) by 11.9%. It outperforms FM on 79% of evaluated breaths, although most patient-grouped comparisons are not significant given the small patient population.

Together, these experiments demonstrate the applicability of PEFM beyond the Windkessel model, with the lowest mean RMSE across all three additional electrophysiological and respiratory settings.

AuthorsOpenReview ↗

W2/Q2, Q3, and References

(W2/Q2) Additional datasets and generalization.

Our original experiments focused on simulated and measured cardiovascular signals, motivated by prior PINN-based studies [3, 4]. To address concerns regarding sample size and generalization, we additionally evaluated PEFM on PWDB [5], containing 3,837 physiologically plausible virtual subjects using the three-element Windkessel model. Subjects span six age groups (25–75 years), heart rates of 62–89 bpm, stroke volumes of 42–82 mL, and systolic pressures of 82–138 mmHg. Each subject contributes one waveform, with subject-disjoint splits of 2,687/575/575 for training/validation/testing.

PWDB results. Test RMSE (mmHg, mean ± std over three seeds) and subject-level paired Wilcoxon pp-values against PEFM, pooled over seeds.

MethodRMSE ↓p-value
Supervised6.756 ± 0.1702×10−182\times10^{-18}
FM8.077 ± 0.1185×10−595\times10^{-59}
WPINN (oracle)6.764 ± 0.15210−2210^{-22}
WPINN (learned)6.728 ± 0.1807×10−167\times10^{-16}
PIFM-GN8.059 ± 0.1442×10−592\times10^{-59}
PBFM (oracle)8.088 ± 0.0919×10−689\times10^{-68}
PBFM (learned)8.165 ± 0.2588×10−658\times10^{-65}
PEFM (learned)6.415 ± 0.108–

PEFM achieves the lowest RMSE, significantly outperforming all seven baselines (subject-level Wilcoxon and paired seed-level tests, p<0.05p<0.05). It reduces RMSE by 20.6% relative to FM and 4.7% relative to WPINN (learned).

The additional datasets also introduce variations in measurement and physiological conditions: CCVW-ICU includes patients aged 35–91 years under different ventilator settings (pressure support 6–12 cmH₂O; PEEP 4–6 cmH₂O), while EDGAR includes distinct electrode configurations (599-electrode pericardial cage and 240-electrode epicardial sock) and activation protocols.

These experiments strengthen evidence for robustness across simulated subjects and diverse experimental conditions. However, EDGAR does not establish inter-heart generalization, and real-world demographic subgroup analyses, particularly BMI-stratified evaluations, remain future work.

(Q3) Downstream clinical evaluation.

Our original manuscript already evaluates clinically relevant SBP and DBP estimation from reconstructed waveforms (Table 3). On BioZGraphene, PEFM achieves the lowest waveform RMSE (14.2 mmHg) and MAE (11.5 mmHg), the smallest absolute SBP mean error (0.61±14.250.61\pm14.25 mmHg), and a competitive DBP error (1.46±11.221.46\pm11.22 mmHg). These findings support the preservation of clinically relevant BP characteristics, beyond waveform reconstruction accuracy. We will emphasize this evaluation in the revised manuscript.

References

[1] K. Aras et al. Experimental Data and Geometric Analysis Repository—EDGAR. Journal of Electrocardiology, 48(6):975–981, 2015. https://doi.org/10.1016/j.jelectrocard.2015.08.008

[2] X. Liu et al. Clinical and ventilator waveforms of Chinese patients during pressure support ventilation. Scientific Data, 13(1):48, 2025. https://doi.org/10.1038/s41597-025-06364-z

[3] K. Sel et al. Physics-informed neural networks for modeling physiological time series for cuffless blood pressure estimation. npj Digital Medicine, 6:110, 2023. https://doi.org/10.1038/s41746-023-00853-4

[4] D. Osman et al. Cardiovascular digital twins using a Windkessel physics informed neural network. npj Digital Medicine, 9:443, 2026. https://doi.org/10.1038/s41746-026-02610-9

[5] P. H. Charlton et al. Modeling arterial pulse waves in healthy aging: a database for in silico evaluation of hemodynamics and pulse wave indexes. American Journal of Physiology-Heart and Circulatory Physiology, 317(5):H1062–H1085, 2019. https://doi.org/10.1152/ajpheart.00218.2019

Reviewer_wZkqOpenReview ↗

Overall Score Raised to 2

Thank you for the additional experiments and detailed explanations. The new results address my concerns regarding Q1 and Q2. I will raise my overall score to 2. Regarding Q3, I agree that estimating systolic and diastolic blood pressure (SBP and DBP) from the reconstructed waveforms is clinically meaningful, and Table 3 reports the accuracy of these estimates.

Could the authors describe a deployment scenario for PEFM and explain how the reconstructed blood pressure waveform would be used in that setting?

We thank the reviewer for this follow-up question. One potential deployment scenario is physiological monitoring in intensive care units (ICUs), as demonstrated by our additional CCVW-ICU experiments. In this setting, PEFM could reconstruct transpulmonary pressure waveforms from continuously measured airway flow using the single-compartment respiratory model. These reconstructed waveforms could help clinicians assess respiratory mechanics, monitor changes in lung compliance and airway resistance, and potentially support ventilator management [1, 2], while reducing reliance on direct esophageal pressure measurements.

Beyond this setting, reconstructed blood pressure waveforms could potentially support downstream hemodynamic analysis. Previous studies have used arterial pressure waveforms to assess pulse pressure and vascular tone [3], estimate stroke volume and cardiac output through pulse wave analysis [4, 5], predict fluid responsiveness [6], and investigate hypotension prediction [7]. These applications highlight the potential value of preserving the full pressure waveform beyond conventional blood pressure measurements.

References

[1] Ball, L., Talmor, D., & Pelosi, P. (2024). “Transpulmonary pressure monitoring in critically ill patients: pros and cons.” Critical Care, 28, 177. https://doi.org/10.1186/s13054-024-04950-y

[2] “Fundamental concepts and the latest evidence for esophageal pressure monitoring.” (2023). Journal of Intensive Care. https://doi.org/10.1186/s40560-023-00671-6

[3] Bertrand, M., et al. (2025). “Applied physiology at the bedside: using invasive blood pressure as a true monitoring tool.” Annals of Intensive Care, 15, 192. https://doi.org/10.1186/s13613-025-01608-y

[4] Thomsen, K. K., Kouz, K., & Saugel, B. (2023). “Pulse wave analysis: basic concepts and clinical application in intensive care medicine.” Current Opinion in Critical Care, 29(3), 215–222. https://doi.org/10.1097/MCC.0000000000001039

[5] Wołos, K., et al. (2024). “Non-invasive assessment of stroke volume and cardiovascular parameters based on peripheral pressure waveform.” PLOS Computational Biology, 20(4), e1012013. https://doi.org/10.1371/journal.pcbi.1012013

[6] Gupta, C. B., et al. (2024). “Improving the precision of shock resuscitation by predicting fluid responsiveness with machine learning and arterial blood pressure waveform data.” Scientific Reports, 14, 2227. https://doi.org/10.1038/s41598-023-50120-5

[7] Yang, H.-L., et al. (2025). “The effect of selection bias on the performance of a deep learning-based intraoperative hypotension prediction model using real-world samples from a publicly available database.” British Journal of Anaesthesia, 135(3), 571–581. https://doi.org/10.1016/j.bja.2025.03.024

Reviewer_kM3pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a physics-informed flow matching framework that imposes the physiological relationship directly on the learned local velocity field through a physical constraint. They derive a velocity-space residual that enforces physical consistency at each flow time. They evaluate translation from virtual aortic flow and from measured bioimpedance to pressure against baselines. Experiment results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. It does not just add PINN loss to the final outpout of flow matching, yet it derives velocity space residual from source-target governing equation. This directly applies to the training objective of flow matching.
  2. The theory analysis is comprehensive, including the error restriction based on physical constraint, the multi-step integration error, and the relation between local physical residual and the final physical residual. It also analyzes how the physical constraint can only fix errors on specific directions.
Critical weaknesses
  1. Velocity space physics loss seens to be reparameterization of endpoint physics. It equals to adding physical loss to the implied endpoint of local velocity. If we predict an implied endpoint for each intermeiate state and add PINN loss, can we get the same result as the velocity space physics loss?
  2. The improvements on these datasets are limited. It's better to include statistical significance test to further verify the effectiveness of the proposed method. More datasets with large sample size are encouraged to be included as well.
Questions for authors

Please check the weakness part.

4 comments

We thank the reviewer for recognizing our velocity-space formulation and theoretical analysis. We (W1) clarify the relationship between velocity-space and endpoint physics regularization and (W2) provide additional experimental evidence demonstrating PEFM's superiority over baselines across cardiovascular, respiratory, and electrophysiological datasets.

(W1) Equivalence to endpoint physics at the implied endpoint.

We agree with the reviewer that both formulations are algebraically equivalent. However, PEFM does not explicitly reconstruct this endpoint. The residual is defined directly on vθv_ \theta, and the implied endpoint is an interpretation rather than a required computation. Our contribution is the derivation of the source-to-target governing relation as local velocity supervision in paired FM, together with its implications for generated trajectories and the resulting error analysis. By enforcing physical consistency at intermediate states, PEFM provides direct physics-based supervision throughout the FM path, which may improve the learned velocity field and its generalization. Moreover, unlike methods that evaluate physics on endpoints obtained through trajectory integration, PEFM avoids computationally expensive rollouts during training.

(W2) Statistical significance and larger datasets.

Prior PINN studies have investigated Windkessel-based blood pressure reconstruction using simulated and bioimpedance signals [1]. To extend our evaluation, we conducted additional experiments on PWDB, EDGAR, and CCVW-ICU, motivated by prior physics-informed applications in cardiovascular hemodynamics [2, 1], inverse electrocardiographic imaging [3, 4], and respiratory mechanics [5]. Results are reported over three seeds with paired statistical tests.

1. PWDB (Arterial haemodynamics) [6]. We evaluated flow-to-pressure translation on 3,837 virtual subjects using the three-element Windkessel model.

PWDB: Test RMSE (mmHg, mean ± std over three seeds) and paired subject-level Wilcoxon pp-values against PEFM, pooled over seeds.

MethodRMSE ↓p-value
Supervised6.756 ± 0.1702×10−182\times10^{-18}
FM8.077 ± 0.1185×10−595\times10^{-59}
WPINN (oracle)6.764 ± 0.15210−2210^{-22}
WPINN (learned)6.728 ± 0.1807×10−167\times10^{-16}
PIFM-GN8.059 ± 0.1442×10−592\times10^{-59}
PBFM (oracle)8.088 ± 0.0919×10−689\times10^{-68}
PBFM (learned)8.165 ± 0.2588×10−658\times10^{-65}
PEFM (learned)6.415 ± 0.108–

PEFM outperforms all seven baselines (p<0.05p<0.05), reducing RMSE by 20.6% relative to FM and 4.7% relative to WPINN (learned). Paired seed-level tests also support this (p<0.05p<0.05).

2. EDGAR (ECGI) [7]. We reconstructed heart-surface potentials from torso measurements using a linear forward operator, evaluating two preparations with 205 and 927 windows.

EDGAR: Test RMSE (mV, mean ± std over three seeds) and grouped Wilcoxon pp-values against PEFM, grouped by beat (u020515) or recording run (u180809), pooled over seeds.

Methodu020515 RMSE ↓u020515 pu180809 RMSE ↓u180809 p
Supervised0.293 ± 0.044<0.0010.757 ± 0.0360.031
FM0.218 ± 0.027<0.0010.721 ± 0.0460.031
WPINN0.306 ± 0.051<0.0010.841 ± 0.0450.031
PIFM-GN0.200 ± 0.0210.1460.646 ± 0.1660.156
PBFM0.217 ± 0.027<0.0010.696 ± 0.1650.156
PEFM0.196 ± 0.025–0.608 ± 0.116–

PEFM achieves the lowest mean RMSE in both preparations, improving on FM by 10.1% and 15.7%. Improvements are significant against four baselines in u020515 (p<0.001p<0.001) and three in u180809 (p=0.031p=0.031), although seed-level significance is not established for u180809.

3. CCVW-ICU (Respiratory mechanics) [8]. We evaluated airway flow-to-transpulmonary-pressure translation on 276 breaths from seven ICU patients using the single-compartment respiratory equation, with patient-disjoint splits.

CCVW-ICU: Test RMSE (cmH₂O, mean ± std over three seeds) and paired patient-grouped Wilcoxon pp-values against PEFM (learned), pooled over seeds.

MethodRMSE ↓p-value
Supervised3.517 ± 0.3630.062
FM3.437 ± 0.6610.062
WPINN (oracle)3.654 ± 0.4270.062
WPINN (learned)2.704 ± 0.8900.312
PIFM-GN3.532 ± 0.7800.062
PBFM (oracle)3.389 ± 0.8640.094
PBFM (learned)3.202 ± 1.0450.031
PEFM (oracle)2.600 ± 0.502–
PEFM (learned)2.382 ± 0.892–

PEFM achieves the lowest mean RMSE, improving on FM by 30.7% and WPINN (learned) by 11.9%.

Overall, PEFM achieves the lowest mean RMSE in all four settings. PWDB establishes strong statistical evidence on a large population, while EDGAR and CCVW-ICU demonstrate improvements on measured physiological data with distinct governing equations. We will incorporate these findings and their statistical limitations into the revised manuscript.

AuthorsOpenReview ↗

References

[1] D. Osman et al. Cardiovascular digital twins using a Windkessel physics informed neural network. npj Digital Medicine, 9:443, 2026. https://doi.org/10.1038/s41746-026-02610-9

[2] K. Sel et al. Physics-informed neural networks for modeling physiological time series for cuffless blood pressure estimation. npj Digital Medicine, 6:110, 2023. https://doi.org/10.1038/s41746-023-00853-4

[3] L. Li et al. Solving the inverse problem of electrocardiography for cardiac digital twins: A survey. IEEE Reviews in Biomedical Engineering, 18:316–336, 2024.

[4] L. Zhu et al. Physics-informed residual learning with spatiotemporal local support for inverse ECG reconstruction. Scientific Reports, 15:31747, 2025. https://doi.org/10.1038/s41598-025-15687-1

[5] A. K. Kumar et al. Physics-informed neural entangled-ladder network for inhalation impedance of the respiratory system. Computer Methods and Programs in Biomedicine, 231:107421, 2023. https://doi.org/10.1016/j.cmpb.2023.107421

[6] P. H. Charlton et al. Modeling arterial pulse waves in healthy aging: a database for in silico evaluation of hemodynamics and pulse wave indexes. American Journal of Physiology-Heart and Circulatory Physiology, 317(5):H1062–H1085, 2019. https://doi.org/10.1152/ajpheart.00218.2019

[7] K. Aras et al. Experimental Data and Geometric Analysis Repository—EDGAR. Journal of Electrocardiology, 48(6):975–981, 2015. https://doi.org/10.1016/j.jelectrocard.2015.08.008

[8] X. Liu et al. Clinical and ventilator waveforms of Chinese patients during pressure support ventilation. Scientific Data, 13(1):48, 2025. https://doi.org/10.1038/s41597-025-06364-z

Dear Reviewer kM3p,

As the discussion phase is approaching its end, we would greatly appreciate it if you could let us know whether our clarifications and additional experiments have adequately addressed your remaining concerns.

We would be happy to address any further questions before the discussion period ends. Thank you again for your time and constructive engagement throughout the discussion!

Reviewer_kM3pOpenReview ↗

Post Rebuttal Comments

I have checked the author's rebuttal, and it has addressed my main concerns. I will keep the positive rating.

Reviewer_Gf72OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a flow matching framework for physiological signal translation that incorporates known physical relationships into the learned velocity field of a flow matching model. Instead of enforcing physical consistency solely on predicted endpoints, the authors derive a velocity-space residual that applies physical constraints at intermediate points along the transport trajectory of the flow matching model. The authors establish a closed-form physics loss and provide a theoretical analysis of how physics regularization and multi-step integration can reduce prediction error and improve physical consistency.

The authors' proposed method is evaluated on simulated aortic flow-to-pressure and measured bioimpedance-to-pressure translation tasks using a two-element Windkessel model. Comparisons against supervised learning, conventional flow matching, and physics-informed baselines show that the proposed methods achieves a good fit on both datasets while maintaining relatively low residuals with respect to the phsyics-based constraints.

Critical strengths

This paper addresses an interesting problem of learning accurate mappings from limited paired data while maintaining consistency with known physical relationships. Incorporating physics into flow matching is a very natural approach that can be potentially useful approach in many applications. The proposed approach is very clearly presented and well motivated, and the paper is technically sound.

Critical weaknesses

(1) The claimed contribution in this submission is to impose the physical constrains directly on the flow matching velocity fields instead of the predicted endpoints. However, for the straight path interpolation in the experiments, isn't the velocity space residual mathematically equivalent to the original physical constraint (applied to a locally reconstructed endpoint), i.e.(\hat{x}1=x_0+v\theta(z_\tau,\tau))? If this is the case, then I think that the proposed physics loss can be interpreted as the same as the conventional physics regularization but applied at intermediate flow matching states, and not a fundamentally new form of physics-based modeling. I understand that the proposed formulation has other advantages in terms of computational complexity (since it avoids trajectory rollouts in training), but the paper does not really sufficiently establish what the conceptual novelty is in this formulation.

(2) Another theoretical concern is that the analysis of endpoint physical consistency is based on controlling the physics residual on the generated trajectory, but the proposed training objective evaluates this residual at intermediate states constructed from ground-truth source-target pairs. Aren't these two trajectoreis generally different when the learned velocity field is imperfect? this would mean that optimizing the proposed training loss may not necessarily guarantee that the physical residual will be small along the trajectory seen during inference. The paper does not provide sufficient assumptions to establish this connection.

The authors can correct me if my understanding is incorrect in either points.

Questions for authors

Please address the weaknesses highlighted above.

2 comments

We thank the reviewer for the careful reading, positive assessment of our work, and constructive comments regarding the conceptual novelty and theoretical justification of our formulation.

(W1) Equivalence to endpoint physics at the implied endpoint.

We agree with the reviewer that the two formulations are algebraically equivalent. For the straight interpolation path, the locally implied endpoint is x^1(v)=x0+vθ(zτ,τ)\hat{x}_ 1^{(v)}=x_ 0+v_ \theta(z_ \tau,\tau). Under the linear governing relation Ax1=Bx0Ax_ 1=Bx_ 0, our velocity residual satisfies

rv(τ)=Avθ(zτ,τ)−(B−A)x0=Ax^1(v)−Bx0.r_ v(\tau)=Av_ \theta(z_ \tau,\tau)-(B-A)x_ 0=A\hat{x}_ 1^{(v)}-Bx_ 0.

Thus, applying the same physics penalty to this locally implied endpoint yields exactly the PEFM loss. However, PEFM does not form this endpoint. The residual is defined directly on vθv_ \theta, and the implied endpoint is an interpretation of it rather than a quantity it requires. Our contribution is the derivation of the source to target governing relation as local velocity supervision in paired flow matching, together with its interpretation for generated trajectories and the resulting error analysis. Unlike methods that evaluate physics after integrating the learned field, PEFM does not require trajectory integration during training.

(W2) Training states versus the generated trajectory.

We agree that the distinction is important. As in standard flow matching, training evaluates the learned field at states constructed from observed pairs, while inference follows the field starting from the source. Errors in predicted velocity may cause the generated trajectory to depart from the reference path. Our existing Proposition 2 establishes that a small physics residual along the generated trajectory controls the physical consistency of the final prediction. However, during training, the VF loss is evaluated along the reference path, and Proposition 2 does not, by itself, establish that minimizing this loss ensures physical consistency at the generated endpoint.

This is why PEFM uses both LFM\mathcal{L}_ {\mathrm{FM}} and LVF\mathcal{L}_ {\mathrm{VF}}. The FM loss encourages the learned velocities to follow the reference path, while the physics loss encourages those velocities to respect the governing relation. Together, they address both the difference between the paths and physical consistency along the path used for training. We can make this connection precise by first bounding the distance between the reference and generated trajectories, and then bounding the physical residual at the generated endpoint.

For a fixed pair (x0,x1)(x_ 0,x_ 1), let zτ=x0+τ(x1−x0)z_ \tau=x_ 0+\tau(x_ 1-x_ 0) denote the reference path and z^τ\hat z_ \tau the path generated by integrating vθv_ \theta from x0x_ 0. Assume that vθv_ \theta is uniformly LL-Lipschitz in its state and that the discretized physical operator AA is bounded. Writing vτ⋆=x1−x0v_ \tau^\star=x_ 1-x_ 0, we obtain

∥z^τ−zτ∥≤∫0τeL(τ−s)∥vθ(zs,s)−vs⋆∥ds.\lVert\hat z_ \tau-z_ \tau\rVert \leq \int_ 0^\tau e^{L(\tau-s)} \lVert v_ \theta(z_ s,s)-v_ s^\star\rVert ds.

This shows that errors in the FM velocity predictions control the distance between the reference and generated paths.

For the straight path, define the reference and generated physics residuals as

rv(τ)=Avθ(zτ,τ)−(B−A)x0,r_ v(\tau)=Av_ \theta(z_ \tau,\tau)-(B-A)x_ 0,

rvtraj(τ)=Avθ(z^τ,τ)−(B−A)x0.r_ v^{\mathrm{traj}}(\tau)=Av_ \theta(\hat z_ \tau,\tau)-(B-A)x_ 0.

The difference between these residuals is controlled by the distance between the paths. Since the endpoint residual equals the integral of rvtrajr_ v^{\mathrm{traj}} along the generated path, combining these two observations gives

∥Az^1−Bx0∥≤ℓVF+∥A∥(eL−1)ℓFM,\lVert A\hat z_ 1-Bx_ 0\rVert \leq \sqrt{\ell_ {\mathrm{VF}}}+\lVert A\rVert (e^L-1)\sqrt{\ell_ {\mathrm{FM}}},

where ℓVF=∫01∥rv(τ)∥2dτ\ell_ {\mathrm{VF}}=\int_ 0^1\lVert r_ v(\tau)\rVert^2 d\tau and ℓFM=∫01∥vθ(zτ,τ)−vτ⋆∥2dτ\ell_ {\mathrm{FM}}=\int_ 0^1\lVert v_ \theta(z_ \tau,\tau)-v_ \tau^\star\rVert^2 d\tau are the physics and FM losses for this pair, whose averages over pairs define LVF\mathcal{L}_ {\mathrm{VF}} and LFM\mathcal{L}_ {\mathrm{FM}}. Thus, under these assumptions, the two losses together control physical consistency at the generated endpoint. This is a bound for a fixed pair under exact integration, rather than an unconditional guarantee for unseen samples. We have included the full proof in a ZIP file as supplementary material and will incorporate it into the camera-ready manuscript.

Dear Reviewer Gf72,

As the discussion phase is approaching its end, we would greatly appreciate it if you could let us know whether our clarifications and additional experiments have adequately addressed your remaining concerns.

We would be happy to address any further questions before the discussion period ends. Thank you again for your time and constructive engagement throughout the discussion!

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Reviewer Concerns and Key Rebuttal Findings

We sincerely thank the AC for overseeing our submission and all reviewers for their constructive feedback.

Below, we summarize the main concerns and key findings from our rebuttal.

To reviewer Gf72

Q1. Equivalence to endpoint physics regularization

We acknowledge that, for straight interpolation paths and linear governing relationships, our velocity-space physics loss is algebraically equivalent to applying physics regularization to the locally implied endpoint. Our contribution is the derivation of this relationship as direct velocity-space supervision, avoiding trajectory integration during training, together with its theoretical analysis.

Q2. Training residuals versus inference trajectories

We derived a new theoretical bound showing that, under stated regularity assumptions, the physics residual at the generated endpoint is controlled jointly by the FM and physics losses evaluated along the reference path. The complete proof is included in our supplementary materials.

To reviewer kM3p

Q3. Equivalence to locally reconstructed endpoint constraints

As clarified in Q1, the formulations are algebraically equivalent under the considered assumptions. PEFM provides a direct velocity-space formulation without requiring trajectory rollouts during training.

Q4. Statistical significance and additional datasets

We conducted additional experiments across three datasets and four settings, covering three distinct governing relationships. PEFM achieves the lowest mean RMSE in all four settings, improving over FM by 20.6% on PWDB, 10.1% and 15.7% on EDGAR, and 30.7% on CCVW-ICU. Improvements are statistically significant against all seven baselines on PWDB and several baselines on EDGAR, while patient-level significance remains limited on CCVW-ICU.

To reviewer wZkq

Q5. Generalization beyond Windkessel modeling

We extended PEFM to ECGI (EDGAR) and respiratory mechanics (CCVW-ICU), demonstrating its applicability to two additional governing relationships. PEFM achieves the lowest mean RMSE in both settings.

Q6. Dataset scale and subject robustness

We expanded our evaluation to 3,837 virtual subjects across six age groups (PWDB), seven ICU patients under varying clinical conditions (CCVW-ICU), and two cardiac preparations with distinct measurement configurations (EDGAR). These experiments strengthen our evaluation across physiological and measurement conditions, while limitations in real-world subject-level generalization remain.

Q7. Downstream clinical relevance

Our original BioZGraphene evaluation already includes SBP and DBP estimation. PEFM achieves the lowest waveform RMSE (14.2 mmHg), MAE (11.5 mmHg), and smallest absolute SBP mean error (0.61 ± 14.25 mmHg), with competitive DBP estimation.

We appreciate the reviewers' feedback and will incorporate these additional experiments, theoretical results, and clarifications into the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.