acceptodds
Under review as a conference paper at ICLR 2027

Physics-Informed Reconstruction of Sparse Hemodynamic Fields

Abstract

Physics-informed neural networks (PINNs) embed governing partial differential equations as constraints within the training loss, but naively weighting this constraint from the start of training traps optimization near a trivial zero-field solution, a premature-physics failure mode that has limited their reliability for sparse-data field recovery. We introduce a training methodology combining a three-phase physics-weight curriculum with exponential-moving-average residual scaling, which defers strong physics enforcement until the network has learned an adequate data-driven representation and adaptively balances multiple governing-equation residuals without manual tuning. We study this methodology on the recovery of complete hemodynamic fields-velocity, pressure, and passive-scalar concentration—from sparse observations, using a CFD-simulated flow within a patient-specific three-dimensional intracranial aneurysm geometry derived from CT imaging, complemented by a synthetic helical-pipe flow with a known analytical solution. Against linear interpolation, proper orthogonal decomposition, and an architecturally identical physics-free network, our approach achieves mean relative error at measurement density - a , , and improvement respectively - decreasing monotonically to at full density. An architecture-matched ablation isolates a further improvement attributable to the embedded governing equations alone, independent of network capacity, while a five-configuration ablation over the curriculum, residual scaling, and gradient clipping confirms the curriculum as the dominant stabilizer. We further probe the method's generalization through robustness to Gaussian measurement noise and sensitivity to spatial acquisition strategy. Throughout, pressure recovery remains directly coupled to the velocity field via the momentum equation, a dynamical consistency unavailable to baselines that fit each field independently. Together, these results show that principled training-schedule design, not architecture alone, is critical to realizing physics-informed learning's promise for sparse biomedical field reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Physics informed neural networks, sparse data field reconstruction, hemodynamic modeling, cerebral aneurysms, physics informed learning

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

This work proposes a new method for reconstruction of sparse hemodynamic fields. They introduce a training methodology based on a three-phase physics-weight curriculum with exponential-moving-average residual scaling. They apply it to the recovery of complete hemodynamic fields from sparse observations. They probe the method's generalization through robustness to Gaussian measurement noise and sensitivity to spatial acuisition strategy. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. The problem of reconstruction hemodynamic fields from sparse observation is important and has practical value for biomedical research and clinical application. The study motivation that PINN is misled by physical loss in early training stage is also clear.
  2. Experimental results show obvious performance improvement of the proposed method compared with baseline methods. They also consider the influence of noise and sampling method, which shows the attention to practical measurement situations.
Critical weaknesses
  1. The methodology contribution seems limited. The main contribution is the training schedule, which is more about the engineer improvements. EMA normalization and gradient clipping are also widely used techniques. This work is more like a cmobination of established methods, instead of a new method.
  2. Comparison with more advanced PINN methods are missing. Only general PINN works are compared in experiments, while more advanced PINN methods are not included. They should be included for a comprehensive comparison to show the advantage of the proposed method.
Questions for authors

Could the authors provide the analysis on the training efficiency, since PINN needs heavy computation whose computational cost might be higher than general neural networks?

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

This paper studies sparse hemodynamic field reconstruction using PINN that predicts velocity, pressure, and passive-scalar concentration from spatiotemporal coordinates. The proposed training strategy combines a three-phase physics-weight curriculum, EMA-based per-equation residual normalization, and gradient clipping. The proposed method was evaluated using a patient-specific intracranial aneurysm CFD dataset and a synthetic helical-pipe flow. The work demonstrated improvements over interpolation, POD, and physics-free neural networks.

Critical strengths
  1. The problem is meaningful, and incorporating governing equations is a well-motivated approach to regularizing reconstruction from sparse observations.

  2. The evaluation includes measurement-density sweeps, noise perturbations, and different spatial sampling strategies. The architecture-matched PureNN comparison distinguishes the benefit of physics-informed training from differences in network capacity.

  3. The component ablations provide useful information about the training procedure. The pressure-unsupervised experiment also helps clarify the limitations of the reconstruction setting.

Critical weaknesses

Limited methodological contribution. The proposed coupling of a physics-weight curriculum and EMA-based residual normalization improves reconstruction in the evaluated setting, and the ablations reveal an interaction between the two components. However, comparisons with existing curriculum-based and adaptive loss-balancing methods are missing, leaving its added value over established PINN training strategies unclear. The interaction is also explained largely qualitatively, without direct evidence for the claimed optimization mechanism. Thus, the paper does not sufficiently establish a methodological advance or generalizable insight beyond an effective combination of existing techniques.

Unconvincing interpretation of pressure recovery. Removing pressure supervision substantially increases pressure error while slightly improving the other fields. This suggests that the physics loss alone may not sufficiently constrain pressure recovery under the proposed training scheme. Selecting checkpoints by data loss may also favor the remaining supervised fields without ensuring accurate pressure reconstruction. The authors instead attribute the failure to gauge ambiguity, although per-snapshot spatial mean subtraction should remove additive pressure offsets. Pressure-gradient errors and momentum residuals should be examined to distinguish gauge effects from limitations in optimization or loss balancing.

Limited validation across flow settings. A single patient-specific geometry and a synthetic pipe provide limited evidence that the proposed training strategy transfers across anatomies or flow conditions. While the within-geometry task is valid, broader claims about reliable PINN training require more diverse validation.

Inconsistencies and missing experimental details. At 10% density, Table 1 reports a mean error of 0.115, whereas the zero-noise per-field values in Section 5.5 average to approximately 0.164, without an explanation of differing settings. The claim of superiority at every density also conflicts with PureNN-6L’s better result at 1%. These discrepancies should be reconciled, and the actual training budgets and hardware should be specified because the curriculum is defined by wall-clock time.

Questions for authors

See Weaknesses

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

The authors use a Navier-Stokes-informed neural network with advection-diffusion modelling for fluid flow reconstruction from 4D flow MRI. They propose a loss term weighting schedule and combine it with online rescaling of the physics residuals.

Critical strengths

While the proposed training curriculum and EMA residual scaling rely on relatively standard optimization heuristics rather than introducing a fundamentally novel algorithmic contribution, the paper's primary strength lies in its empirical execution and ablation studies. In particular, the systematic evaluation across measurement densities, noise levels, and sensor placement strategies provides practical insights into the training dynamics and interaction effects of physics-informed neural networks for hemodynamic flow reconstruction.

Critical weaknesses

The discussion of related works is lackluster and it is possible that the authors are not sufficiently familiar with important papers in the context of their application.

The proposed training strategy consists of incremental adaptations and does not warrant a methodological contribution in my opinion. However, the performed ablations are interesting and provide insight into physics-informed neural networks in general. I encourage the authors to submit their work to an application-focused venue.

Questions for authors

1 Introduction: I would not consider most of these bullet points as contributions, only the first and maybe second (if you publish the data).

2 Related work:

  • I am missing related work on 4D flow super-resolution, e.g., Fathi et al. "Super-resolution and denoising of 4D-Flow MRI using physics-Informed deep neural nets" (2020), Kissas et al. "Machine learning in cardiovascular flows modeling: Predicting arterial blood pressure from non-invasive 4D flow MRI data using physics-informed neural networks" (2020)
  • More importantly, given the authors' claim about their contribution, please discuss Liu et al. "ConFIG: Towards Conflict-free Training of Physics Informed Neural Networks" (ICLR 2025) and Datar et al. "Fast training of accurate physics-informed neural networks without gradient descent" (ICLR 2026)

3.1 Problem formulation and network: What does it mean for a field to be kept O(1)? If you mean that they are roughly of the same magnitude, I am not sure this is the right notation.

3.2 The premature-physics failure mode: What do you mean by "the convective and diffusive terms in Equations 1–4"? Do equations (2) to (4) not describe momentum?

4 Experimental setup: - What are "unfillable extrapolated points" in the context of interpolation? - Please elaborate on baseline (ii). How is this used for super-resolution? - What is the ground truth in your metrics? Is it numerical simulation results? I might have missed it, please state it here.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.