pith. sign in

arxiv: 2606.00349 · v1 · pith:PO7JSJAHnew · submitted 2026-05-29 · 💻 cs.LG · cs.AI· cs.CE

(HB-ARFM) History-Bootstrapped Flow Matching for Inverse Boiling Reconstruction

Xianwei Zou , Sheikh Md Shakeel Hassan , Arthur Feeney , Aparna Chandramowlishwaran This is my paper

Pith reviewed 2026-06-28 22:52 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.CE
keywords flow matchinginverse reconstructionboiling dynamicsspatiotemporal fieldspartial observabilityautoregressive modelinghistory bootstrappingfluid dynamics
0
0 comments X

The pith

History-bootstrapped autoregressive flow matching reconstructs valid boiling velocity and temperature fields from partial interface observations.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper addresses reconstructing full spatiotemporal fields such as velocity and temperature from partial observations in boiling, where incomplete data turns even Markovian dynamics into a non-Markovian inverse problem. It introduces HB-ARFM, which first applies conditional flow matching on observation history to bootstrap an initial reconstruction that reduces ambiguities, then uses the same model autoregressively while conditioning on both fresh observations and previous predictions to advance the solution in time. Evaluation on two inverse tasks with different sparsity levels shows the approach yields physically consistent and temporally coherent full fields where other models produce invalid results. Readers would care because many scientific problems, from fluid imaging to atmospheric inference, require recovering hidden states from sparse measurements without direct access to the complete dynamics.

Core claim

HB-ARFM produces physically and temporally valid reconstructions of full velocity and temperature fields from interface geometry and motion in boiling dynamics by first using conditional flow matching on observation history for bootstrapping and then applying the model autoregressively on new observations and past predictions, succeeding across two inverse tasks with different observation sparsity where other models fail.

What carries the argument

History-bootstrapped autoregressive flow matching (HB-ARFM), which conditions on observation history to bootstrap the initial reconstruction via conditional flow matching and then propagates forward autoregressively while conditioning on both new observations and prior predictions.

If this is right

  • Recovers complete velocity and temperature fields from interface geometry and motion alone.
  • Maintains physical and temporal validity across tasks that differ in observation sparsity.
  • Avoids the invalid outputs produced by models lacking history bootstrapping or autoregressive conditioning.
  • Resolves the ill-posed inverse problem created when partial observation operators make the posterior non-Markovian.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same bootstrapping-plus-autoregression pattern may apply directly to other partial-observation inverse problems such as atmospheric state recovery from satellite imagery.
  • History conditioning appears necessary whenever the observation operator itself introduces temporal dependencies that single-timestep models cannot capture.
  • Longer test sequences would expose whether autoregressive propagation eventually diverges even when initial bootstrapping succeeds.

Load-bearing premise

Conditioning on observation history plus autoregressive use of past predictions is sufficient to resolve the non-Markovian posterior induced by partial observations without accumulating errors over time.

What would settle it

Running the method on a held-out sequence with known full ground-truth states and checking whether reconstructed velocity and temperature fields remain consistent with the boiling PDE dynamics and match ground truth over many timesteps beyond the training horizon.

Figures

Figures reproduced from arXiv: 2606.00349 by Aparna Chandramowlishwaran, Arthur Feeney, Sheikh Md Shakeel Hassan, Xianwei Zou.

Figure 1
Figure 1. Figure 1: An Example Pipeline of Spatiotemporal Boiling Dynamics Reconstruction from High-speed Imaging. Sequential high-speed images are processed through segmentation and optical flow to extract phase field and interface velocity. These extracted observations condition our proposed HB-ARFM model which then outputs full spatiotemporal predictions of velocity and temperature fields, enabling complete multiphase flui… view at source ↗
Figure 2
Figure 2. Figure 2: History-Bootstrapped ARFM. The model combines a history encoder processing temporal sequences with an FM UNet for both initial and AR reconstructions. The FM UNet predicts the full spatiotemporal fields. The green path shows initial reconstruction while the red feedback loop enables sequential reconstruction with data assimilation. Algorithm 2 Inference: History-Bootstrapped ARFM Require: Observations y0:T… view at source ↗
Figure 3
Figure 3. Figure 3: Snapshot Reconstruction of Temperature and Velocity Fields in Subcooled Pool Boiling. Given SDF and interface velocity as input, models predict full temperature (top) and velocity (bottom) fields. Generative methods reconstruct near-interface regions well but differ in bulk field predictions. HB-ARFM produces sharper gradients and more physically consistent reconstructions. VE-SDE exhibits catastrophic fai… view at source ↗
Figure 4
Figure 4. Figure 4: HB-ARFM History-length Ablation. Shaded regions denote ±3σ over 30 random seeds. 5.2. Temporal Consistency and Rollout Reconstruction [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Rollout of Reconstructed Temperature Fields in Subcooled Pool Boiling. HB-ARFM produces temporally consistent dynamics across timesteps by combining reconstruction with data assimilation in an autoregressive framework. HistoryFM and DiffusionPDE exhibits inconsistency between timesteps. Deterministic baselines (Bubbleformer and UNet) produce blurred reconstructions or diverge immediately when initialized f… view at source ↗
Figure 6
Figure 6. Figure 6: Physics Validation of HB-ARFM. Left: Spatial temperature profiles at different wall-normal distances (y = 0, 8, 16 in non-dimensional units). The y = 0 profile corresponds to the heated surface where the thermal boundary layer originates. The heater extends in X position from 24 to 104 in pixel space. Right: Wall heat flux as a function of wall temperature Twall evaluated over 300 timesteps. Training data … view at source ↗
Figure 7
Figure 7. Figure 7: HB-ARFM Reconstruction Results for Flow Boiling. HB-ARFM reconstructs coupled temperature and velocity fields from sparse interface observations in flow-boiling dominated by strong streamwise advection. The model preserves near-wall thermal boundary layers, downstream wake structures, and velocity fields consistent with the evolving interface dynamics. den fields. Others such as Denoising Diffusion Operato… view at source ↗
Figure 8
Figure 8. Figure 8: Temperature Prediction from SDF (Task 1) in Subcooled Pool Boiling. Given the signed distance function describing bubble position, models reconstruct the temperature field. The colorbar indicates temperature in ◦C [PITH_FULL_IMAGE:figures/full_fig_p020_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Rollout Metrics for Subcooled Pool Boiling over 300 timesteps. 25 [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Ablation for Different Initial States. HB-ARFM (Zeros): Feeds a completely blank initial state. HB-ARFM (Mean Cond.): Averages the observation history over 10 steps [PITH_FULL_IMAGE:figures/full_fig_p026_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Adding Noise on Observations during Inference. 26 [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
read the original abstract

Reconstructing spatiotemporal fields from partial observations is fundamental to scientific inference, from inferring atmospheric states from satellite data to recovering fluid states from imaging. When observations are incomplete, the inverse problem is fundamentally ill-posed: even when the underlying PDE dynamics are Markovian in the full state, partial observation operators induce a non-Markovian posterior that cannot be resolved from a single timestep. We propose a history-bootstrapped autoregressive flow matching (HB-ARFM) for spatiotemporal inverse reconstruction under partial observability. Observation history bootstraps the initial reconstruction via conditional flow matching, reducing ambiguities. The same conditional transport model is then applied autoregressively, conditioning on both new observations and past predictions to propagate the reconstruction forward in time. We evaluate the method on boiling dynamics reconstruction, recovering full velocity and temperature fields from interface geometry and motion. Across two inverse tasks with varying observation sparsity, HB-ARFM produces physically and temporally valid reconstructions where other models fail.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes History-Bootstrapped Autoregressive Flow Matching (HB-ARFM) to reconstruct spatiotemporal fields (velocity and temperature) from partial observations in boiling dynamics. Observation history is used to bootstrap an initial reconstruction via conditional flow matching; the same model is then applied autoregressively, conditioning on new observations plus prior predictions, to address the non-Markovian posterior induced by incomplete observations. The paper claims that this produces physically and temporally valid reconstructions on two inverse tasks with varying sparsity, where other models fail.

Significance. If the central claims hold with supporting quantitative evidence, the work could contribute a practical approach to ill-posed inverse problems in fluid dynamics by combining history conditioning with autoregressive flow matching to maintain temporal consistency under partial observability.

major comments (2)
  1. [Abstract] Abstract: the claim that autoregressive application of the conditional transport model 'resolves the non-Markovian posterior' and yields 'physically and temporally valid reconstructions' is load-bearing, yet the description provides no indication of mitigation for exposure bias (conditioning on model predictions rather than ground truth at test time), such as scheduled sampling, consistency regularization, or bounded-error analysis; without this, error compounding over timesteps remains an unaddressed risk to the temporal-validity claim.
  2. [Abstract] Abstract: the assertion of success 'across two inverse tasks with varying observation sparsity' and that 'HB-ARFM produces ... reconstructions where other models fail' is unsupported by any equations, dataset details, baseline methods, quantitative metrics, or ablation results, preventing verification that the architecture actually delivers the stated performance.
minor comments (1)
  1. [Title] The parenthetical '(HB-ARFM)' in the title is redundant with the expansion given in the first sentence of the abstract.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed feedback on our manuscript. We address the major comments point by point below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that autoregressive application of the conditional transport model 'resolves the non-Markovian posterior' and yields 'physically and temporally valid reconstructions' is load-bearing, yet the description provides no indication of mitigation for exposure bias (conditioning on model predictions rather than ground truth at test time), such as scheduled sampling, consistency regularization, or bounded-error analysis; without this, error compounding over timesteps remains an unaddressed risk to the temporal-validity claim.

    Authors: We acknowledge the referee's concern regarding exposure bias in the autoregressive rollout. The manuscript describes conditioning on both new observations and prior predictions to propagate reconstructions, which is intended to leverage the history bootstrapping to mitigate ambiguities from partial observations. However, the abstract does not explicitly discuss mitigation techniques such as scheduled sampling or error bounds. We will revise the manuscript to add a dedicated discussion of this issue, including any empirical checks on error accumulation performed during evaluation. revision: yes

  2. Referee: [Abstract] Abstract: the assertion of success 'across two inverse tasks with varying observation sparsity' and that 'HB-ARFM produces ... reconstructions where other models fail' is unsupported by any equations, dataset details, baseline methods, quantitative metrics, or ablation results, preventing verification that the architecture actually delivers the stated performance.

    Authors: The abstract is a concise summary of the work. The full manuscript provides the model equations for the conditional flow matching, details on the two boiling dynamics inverse tasks and observation sparsity levels, descriptions of baseline methods, quantitative metrics (e.g., field reconstruction errors), and ablation studies demonstrating where HB-ARFM succeeds while others fail. These elements are presented in the methods, experiments, and results sections to support the claims. revision: no

Circularity Check

0 steps flagged

No circularity: method is a standard conditional flow-matching architecture with autoregressive rollout

full rationale

The provided abstract and description define HB-ARFM as a conditional flow-matching model trained on observation histories, then applied autoregressively. No equations, fitted parameters, or self-citations are shown that reduce any claimed result to its inputs by construction. The central mechanism (history bootstrapping + autoregressive conditioning) is an architectural choice whose validity is evaluated externally on boiling reconstruction tasks; it does not rename or tautologically reproduce any input quantity. This matches the default expectation of a non-circular proposal.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only abstract available; ledger reflects explicitly stated premises in the provided text.

axioms (1)
  • domain assumption Partial observation operators induce a non-Markovian posterior that cannot be resolved from a single timestep even when full-state PDE dynamics are Markovian.
    Directly stated in the abstract as the fundamental reason a history-based approach is needed.

pith-pipeline@v0.9.1-grok · 5711 in / 1098 out tokens · 21097 ms · 2026-06-28T22:52:01.669291+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 4 canonical work pages

  1. [1]

    Bastek, J.-H., Sun, W., and Kochmann, D

    URL https://openreview.net/forum? id=tAf1KI3d4X. Bastek, J.-H., Sun, W., and Kochmann, D. Physics-informed diffusion models. InThe Thirteenth International Confer- ence on Learning Representations, 2025. URL https: //openreview.net/forum?id=tpYeermigp. Bertsch, S. S., Groll, E. A., and Garimella, S. V . A compos- ite heat transfer correlation for saturate...

    work page doi:10.1016/j.ijheatmasstransfer.2023.124656 2025

  2. [2]

    Austin Harris, Tom Klosterman, Rajeev Jain, Johann Rudi, Bronson Messer, Michael Pajkos, Jared Carlson, Ran Chu, Mohamed Wahib, Saurabh Chawdhary, Paul M

    doi: https://doi.org/10.1016/j.softx.2022.101168. URL https://www.sciencedirect.com/ science/article/pii/S2352711022001030. El-Gazzar, A. and van Gerven, M. Probabilistic forecast- ing via autoregressive flow matching.arXiv preprint arXiv:2503.10375, 2025. Fu, Q., Suh, Y ., Zhang, X., Chang, S., and Won, Y . Bub- ble2heat: Optical to thermal inference in ...

    work page doi:10.1016/j.softx.2022.101168 2022

  3. [3]

    Konishi, C., Lee, H., Mudawar, I., Hasan, M

    URL https://www.sciencedirect.com/ science/article/pii/S001793101400934X. Konishi, C., Lee, H., Mudawar, I., Hasan, M. M., Nahra, H. K., Hall, N. R., Wagner, J. D., May, R. L., and Mackey, J. R. Flow boiling in microgravity: Part 2 – critical heat flux interfacial behavior, experimental data, and model.International Journal of Heat and Mass Transfer, 81:7...

    work page doi:10.1016/j.ijheatmasstransfer.2014.10 2014

  4. [4]

    Li, Z., Liu-Schiaffini, M., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandku- mar, A

    URL https://www.sciencedirect.com/ science/article/pii/S0017931014009375. Li, Z., Liu-Schiaffini, M., Kovachki, N., Azizzadenesheli, K., Liu, B., Bhattacharya, K., Stuart, A., and Anandku- mar, A. Learning chaotic dynamics in dissipative systems. Advances in Neural Information Processing Systems, 35: 16768–16781, 2022. Li, Z., Han, W., Zhang, Y ., Fu, Q.,...

    2022

  5. [5]

    org/CorpusID:122454757

    URL https://api.semanticscholar. org/CorpusID:122454757. Ni, S., Pan, C., Hibiki, T., and Zhao, J. Applications of nucleate boiling in renewable energy and thermal man- agement and recent advances in modeling——a review. Energy, 289:129962, 2024. 11 History-Bootstrapped Flow Matching for Inverse Boiling Reconstruction Phillips, B. A.Experimental investigat...

    Pith/arXiv arXiv 2024

  6. [6]

    Song, Y ., Shen, L., Xing, L., and Ermon, S

    URL https://openreview.net/forum? id=PxTIG12RRHS. Song, Y ., Shen, L., Xing, L., and Ermon, S. Solving inverse problems in medical imaging with score-based genera- tive models. InInternational Conference on Learning Representations, 2022. URL https://openreview. net/forum?id=vaRCHVj0uGI. Suh, Y ., Chandramowlishwaran, A., and Won, Y . Recent progress of a...

    2022

  7. [7]

    Wang, S., Dou, Z., Shan, S., Liu, T.-R., and Lu, L

    URL https://openreview.net/forum? id=tmIiMPl4IPa. Wang, S., Dou, Z., Shan, S., Liu, T.-R., and Lu, L. Fundiff: Diffusion models over function spaces for physics-informed generative modeling.arXiv preprint arXiv:2506.07902, 2025a. Wang, S., Wu, Z., van Dijk, D., and Lu, L. Geofunflow: Geometric function flow matching for inverse opera- tor learning over co...

    work page doi:10.3390/sym12050721 2025

  8. [8]

    org/CorpusID:121594079

    URL https://api.semanticscholar. org/CorpusID:121594079. 12 History-Bootstrapped Flow Matching for Inverse Boiling Reconstruction A. Software and Data Code, trained models, and evaluation scripts are available at:https://github.com/therml-ai/HB-ARFM. B. Notation Table 2.Notation summary for generative models Symbol Description Xt Full physical state at ti...

  9. [9]

    During training, a clean sample x0 is progressively corrupted by noise ϵ according to a fixed schedule, producing noisy versions xt at different timesteps

    Denoising Diffusion Probabilistic Models (DDPM)DDPM (Ho et al., 2020) defines a forward noising process and learns to reverse it by predicting Gaussian noise at each timestep. During training, a clean sample x0 is progressively corrupted by noise ϵ according to a fixed schedule, producing noisy versions xt at different timesteps. A neural network is train...

    2020

  10. [10]

    Score-based models learn the score function ∇x logp t(x), the gradient of the log-density of the noisy data at each noise level

    Variance Exploding Score-Based Models (VE-SDE).VE-SDE is a score-based generative baseline (Song et al., 2021). Score-based models learn the score function ∇x logp t(x), the gradient of the log-density of the noisy data at each noise level. Training perturbs clean data with a forward stochastic differential equation (SDE) and matches the network output to...

    2021

  11. [11]

    Flow Matching.Flow Matching (Lipman et al., 2023) is a framework for learning continuous-time generative models that transport samples from a simple base distribution to a target distribution by integrating a learned velocity field. Given conditioning information c, flow matching learns a time-dependent velocity fieldvθ(x,c, s) defining the ordinary diffe...

    2023

  12. [12]

    It uses spatial and temporal attention to capture long-range spatial correlations and temporal dependencies arising from phase-change dynamics

    Bubbleformer.Bubbleformer (Hassan et al., 2025) is a spatiotemporal transformer-based neural surrogate for multiphase boiling simulations that can act as a robust forecasting model for a boiling system. It uses spatial and temporal attention to capture long-range spatial correlations and temporal dependencies arising from phase-change dynamics. The model ...

    2025

  13. [13]

    FFNO.FFNO (Tran et al., 2023) is a 2D Factorized Fourier Neural Operator: it applies learnable spectral convolutions in the Fourier domain along each spatial dimension separately, retaining a fixed number of low-frequency modes per dimension, then passes the result through small feedforward networks. Our implementation uses shared Fourier weights and shar...

    2023

  14. [14]

    DiffusionPDE.DiffusionPDE (Huang et al., 2024) is a conditional diffusion baseline for PDE field prediction under partial observation. It uses the EDM (Karras et al., 2022) framework: a denoising network is trained to predict the clean target given noisy target and conditioning, with input/output preconditioning (scale and skip terms depending on noise le...

    2024

  15. [15]

    UNet.UNet (Ronneberger et al., 2015) is a standard convolutional encoder–decoder baseline. The 2D UNet uses a four-level encoder with repeated blocks of two 3×3 convolutions, batch normalization, and GELU activation, downsampling via max pooling; a bottleneck at the deepest level; and a symmetric decoder with transposed convolutions for upsampling and ski...

    2015

  16. [16]

    PDEDiff.PDEDiff (Shysheya et al., 2024) is a conditional diffusion model that generates a full temporal window of w frames jointly rather than predicting one frame at a time. The window is flattened along the channel axis,X∈R w·Cout×H×W , and the model is trained with an amortized mask-based conditioning scheme: the input to the score network is the conca...

    2024

  17. [17]

    Given a window of past conditioning frames {(SDFt−w+1,u Γ t−w+1),

    HistoryFM.History-window Flow Matching is a history-conditioned extension of OT-CFM (described above in §3) that augments the per-frame conditioning with a sliding window of w past observable frames, providing the model with explicit temporal context while remaining non-autoregressive. Given a window of past conditioning frames {(SDFt−w+1,u Γ t−w+1), . . ...

    arXiv 2025