Pith. sign in

REVIEW 3 major objections 5 minor 62 references

Bubbleformer: Forecasting Boiling with Transformers

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Bubbleformer is the first neural model to forecast boiling autonomously, including renucleation, from past fields alone.

desk verdict A valuable dataset and a harder forecasting task, but the 'autonomous nucleation' claim is confounded by conditioning on the nucleation wait time, and the SOTA claims are overstated. read the letter →

arxiv 2507.21244 v1 pith:BGTIO6JT submitted 2025-07-28 cs.LG cs.AIcs.CE

Sheikh Md Shakeel Hassan , Xianwei Zou , Akash Dhruv , Vishwanath Ganesan , Aparna Chandramowlishwaran This is my paper
classification cs.LGcs.AIcs.CE
keywords boilingforecastingspatiotemporaltransformersbubblenucleationneuralPDEsurrogatesflowpoolphysics-basedmetricsML2.0
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

Boiling is chaotic, stochastic, and multiscale, which makes it a hard test for neural PDE surrogates. Existing models cannot forecast boiling because they need future bubble positions as input and fail to learn nucleation from past states; they also fail on flow boiling velocity fields. Bubbleformer is a transformer-based spatiotemporal model that claims to be the first to forecast full boiling dynamics—temperature, velocity, and interface fields—autonomously, including the appearance of new bubbles, without any simulation data at inference. If the claim holds, it would make real-time forecasting, parametric studies, and design exploration of boiling systems practical, since a trained model could replace days of petascale simulation with a fast rollout.

What carries the argument

The load-bearing mechanism is a spatiotemporal transformer whose blocks combine factorized space-time attention with axial attention: temporal attention first along the time axis, then two one-dimensional attentions along height and width, reducing complexity from $\mathcal{O}((HWT)^2)$ to $\mathcal{O}(H^2+W^2+T^2)$ while keeping a global receptive field and a directional bias useful for flow boiling. Frequency-aware attention and feature scaling act as an adaptive sharpening filter to preserve high-frequency interface detail. A FiLM layer conditions every patch embedding on a nine-dimensional thermophysical descriptor (Reynolds, Prandtl, and Stefan numbers; relative viscosity, density, conductivity, and heat capacity; heater temperature; and nucleation wait time), which is what lets one model span fluids and operating conditions.

What would settle it

Train Bubbleformer without the nucleation wait time in its conditioning descriptor (or with the wait time fixed to one constant) and run a 200-step autoregressive rollout; if new bubbles stop appearing at nucleation sites or the Eikonal and vapor-volume metrics degrade sharply, the claim of learned, autonomous nucleation is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that Bubbleformer, a transformer-based spatiotemporal model, is the first model to demonstrate autonomous, physically plausible forecasting of boiling dynamics. Given a history of signed-distance (interface), temperature, and velocity fields, the model predicts future frames of all three fields, and over long autoregressive rollouts it reproduces renucleation events, keeps the predicted heat-flux distribution close to the simulation's, maintains near-unit Eikonal residuals for the signed distance field, and conserves vapor volume. The paper further claims generalization across three fluid classes (dielectric, refrigerant, cryogen), pool and flow boiling, and bubbly/slug/annular regimes, and shows state-of-the-art accuracy on the prediction task when compared with UNet and factorized Fourier neural operator baselines.

Load-bearing premise

The model's renucleation behavior may simply be read off the nucleation wait time fed into its conditioning layer, so the paper's claim that it learns to create new bubbles from past states depends on an ablation that is not performed.

Editorial extensions

If this is right

  • Autoregressive forecasting becomes possible for boiling: a trained model can roll out temperature, velocity, and interface fields for hundreds to thousands of steps without future bubble positions or simulation data.
  • A single conditioned model can generalize across working fluids (FC-72, R515B, LN2), boiling configurations, and flow regimes, reducing the need for per-fluid retraining.
  • The proposed physics-based metrics—heat-flux distribution KL divergence, Eikonal loss, and relative vapor volume error—give a way to check physical plausibility of chaotic forecasts beyond pixel error.
  • BubbleML 2.0 provides over 160 simulations across fluids and regimes, establishing a benchmark for forecasting and prediction of two-phase boiling.
  • Predictions on subcooled pool boiling remain stable over an 800-step rollout (and up to 2000 steps), while UNet and FNO baselines diverge.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the nucleation wait time were removed from the FiLM descriptor and renucleation persisted, that would confirm the model learns the nucleation process from spatiotemporal history rather than reading it off a conditioning parameter; the paper does not report this ablation, so the autonomous-nucleation claim rests on that untested premise.
  • The same physics-based evaluation recipe could transfer to other chaotic multiphase surrogates, such as condensation or cavitation, where aggregate field errors are uninformative but heat-flux statistics, level-set geometry, and mass conservation are the quantities an engineer actually cares about.
  • Extending the architecture to operate directly on adaptive mesh refinement grids—flagged as a limitation by the authors—would let the approach handle water and realistically sized domains, since interpolating AMR data to uniform grids currently injects error.
  • A practical deployment would need to estimate the nucleation wait time online from past frames or marginalize over it, because a real operator will not usually know this stochastic simulation parameter in advance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper presents Bubbleformer, a transformer-based spatiotemporal model that takes a history of temperature, velocity, and signed-distance fields and predicts future fields, including renucleation events, without requiring future bubble positions at inference. The authors introduce BubbleML 2.0, a dataset of over 160 Flash-X simulations across three fluids, pool and flow boiling configurations, and multiple flow regimes. They also propose three physics-based evaluation metrics: heat-flux distribution KL divergence, Eikonal loss for interface geometry, and relative vapor volume error. The headline claims are that Bubbleformer is the first model to demonstrate autonomous, physically plausible forecasting of boiling dynamics, and that it sets state-of-the-art results on both prediction and forecasting benchmarks.

Significance. If the central claims hold, this is a substantial contribution: an architecture that maintains stable autoregressive rollouts in a chaotic multiphase setting, a large open dataset with diverse fluids and regimes, and evaluation metrics that go beyond pixel-wise error are all valuable to the neural-PDE-surrogate community. The paper also ships code, model weights, and dataset DOIs, which supports reproducibility. However, the load-bearing claim of autonomously learned nucleation rests on a conditioning input that may directly supply the renucleation schedule, and the quantitative evidence in the paper contains inconsistencies with the stated state-of-the-art and physical-fidelity claims. The significance is therefore conditional on resolving these points.

major comments (3)
  1. [Section 4.1, Appendix C.1, and Appendices B.4-B.5] The claim that Bubbleformer 'learns nucleation' from past states is confounded by the FiLM conditioning input. The 9-dimensional descriptor used by the FiLM layer (Section 4.1, Table 4, Appendix C.1) includes 'nucleation wait time,' which is the simulation parameter t*_wait that, according to the nucleation model in Appendices B.4 and B.5, directly prescribes how long after bubble departure a new bubble is inserted at each nucleation site. If the model uses this scalar as a clock, then it is being told the renucleation schedule rather than inferring it from the field history. The paper contains no ablation that removes this input, and no analysis of whether predicted nucleation events align with or diverge from the supplied wait time. Without such an ablation, the central claim that Bubbleformer 'forecasts boiling dynamics including nucleation ... without dependence on simulation data during inference' (Abstract) is not supported. A concrete remedy would be to retrain with the wait time removed from the FiLM descriptor and to compare nucleation timing statistics against the ground-truth schedule.
  2. [Section 6.3 and Table 10] The statement that Bubbleformer achieves 'state-of-the-art accuracy across all reported metrics' is contradicted by the authors' own Table 10. On the subcooled pool boiling (FC-72) prediction task, FFNO reports lower MaxErr than Bubbleformer for temperature (2.488 vs. 3.676), for velocity X (0.348 vs. 0.785), and for velocity Y (0.792 vs. 1.700); FFNO also reports lower BRMSE and IRMSE on both velocity components, and lower high-frequency Fourier errors on Velocity X and Velocity Y. The claim should be revised to state where Bubbleformer is and is not state-of-the-art, or additional analysis should be provided to reconcile these discrepancies. In addition, Tables 11 and 12 report only Bubbleformer results; without baseline comparisons, the 'benchmark' claims for saturated pool boiling and flow boiling prediction are not substantiated.
  3. [Section 6.2 with Tables 6 and 8] The physical-fidelity claims are not supported by the reported forecasting metrics. Section 6.2 states that for a 200-step flow boiling rollout the model's 'mass conservation is closely followed,' and Section 4.2 suggests Eikonal losses below 0.1 are sufficient for stable interface evolution. However, Table 8 reports relative vapor volume errors of 2.3 to 3.6 for Bubbleformer-S on subcooled pool boiling, meaning the predicted vapor volume deviates from ground truth by 230-360%, and Table 6 reports Eikonal losses of 0.15-0.18 and KL divergences up to 1.06 for FC-72 single-bubble forecasting. These numbers are difficult to reconcile with 'physically plausible forecasting' and 'closely followed' mass conservation. The authors should either report which specific configurations support the physical-plausibility claim, or temper the claim to reflect the large errors observed on several test cases.
minor comments (5)
  1. [Throughout] There are multiple typos and inconsistent spellings, for example 'domenstrates' in Section 6.3, 'Hieararchical' in Appendix C.1, and 'balance f forces' in Appendix A; these should be corrected.
  2. [Appendix C.3] The text says the training set consists of 12 simulations but then states that 2 out-of-distribution and 3 in-distribution trajectories are left out, which appears to count 17 simulations total; the counting should be clarified.
  3. [Tables 7-9] The forecasting tables report Bubbleformer-S and Bubbleformer-L, but the text in Appendix C.5 says 'we observe that the Bubbleformer-S model performs better than the Bubbleformer-L model across all three metrics'; this is consistent with Table 9, but the abstract's claim of setting 'new benchmark results' is not reflected by any baseline comparison in the forecasting tables, so it would be helpful to restate this as a limitation.
  4. [Section 4.2] The Eikonal loss threshold of 0.1 is introduced as a practical observation, but no sensitivity analysis is provided; a sentence explaining how this threshold was chosen would improve reproducibility.
  5. [Appendix D.2] The interface RMSE definition uses a set I of interface pixels, but the paper does not specify how I is derived from the signed distance field; a precise definition (e.g., pixels within a band |phi| < epsilon) is needed for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

Nucleation-autonomy claim is confounded: the FiLM conditioner receives the simulation's re-nucleation wait time, with no ablation removing it.

  1. other [Section 4.1 and Appendix C.1 (FiLM conditioning inputs), Section 6.2 (renucleation claim), Appendix B.4 (nucleation model)]
    "At each timestep, the model checks whether the four cells surrounding a nucleation site are filled with liquid. If so, the site is marked for re-nucleation after a specified waiting time t∗wait. [...] FiLM Layer 9 fluid parameters are used to condition the model. These parameters are as follows: Reynolds Number(Re), relative specific heat capacity(C′p), relative viscosity(µ′), relative density(ρ′), relative thermal conductivity(k′), Stefan Number(St), Prandtl Number(Pr), nucleation wait time and heater temperature."

    In the simulator, renucleation is not learned but scheduled: a site whose four surrounding cells are liquid is re-nucleated after the prescribed t∗wait (B.4), and t∗wait is computed a priori from heat-flux partitioning (B.5.1). Bubbleformer is given this exact t∗wait as one of its nine FiLM conditioning inputs (C.1). Consequently, the paper's claim that 'Bubbleformer successfully learns this behavior' and the abstract's claim of 'autonomous ... forecasting ... including nucleation' are confounded: the model can reproduce renucleation timing by reading the supplied scalar, without deriving the nucleation schedule from past fields.

full rationale

One circular/confounded step is identified: the nucleation-autonomy headline is undermined by the FiLM wait-time input, which directly prescribes each renucleation in the simulator. The remainder of the paper is not circular: Bubbleformer is trained on held-out simulations from BubbleML 2.0, compared against UNet and FFNO baselines from prior work, and evaluated with independent physics-based metrics (heat-flux KL divergence, Eikonal residual, vapor volume). Self-citations to BubbleML are used for task definitions and baselines rather than to justify Bubbleformer's central mechanism, so they are not load-bearing. The 'state-of-the-art accuracy across all reported metrics' statement is contradicted by the paper's own Table 10 on MaxErr, velocity IRMSE, and high-frequency Fourier errors, but that is an accuracy overclaim rather than circularity. The score reflects the single central-claim confound; the velocity, temperature, and interface forecasting results retain independent content.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claims rest on three categories of uncharged inputs: hand-chosen simulation parameters (nucleation wait time, contact angle limiting velocity), the physical fidelity of the Flash-X simulation data (only qualitatively validated), and the preservation of physics under AMR-to-regular-grid interpolation. No new physical entities are introduced.

free parameters (3)
  • nucleation wait time per fluid = 0.4 tc (FC-72), 0.6 tc (R515B), 1.0 tc (LN2)
    Chosen by hand to set renucleation frequency in simulations and used as a FiLM conditioning input, so the model's nucleation behavior is partly determined by this parameter.
  • ulim (contact angle limiting velocity) = 20-25% of characteristic velocity uc
    Hand-tuned simulation parameter in the contact line model (Appendix B.3); influences generated data.
  • Eikonal loss threshold = 0.1
    Ad hoc threshold claimed sufficient for stable autoregressive rollout (Section 4.2); used as a stability criterion.
assumptions (3)
  • domain assumption Flash-X simulations are a faithful representation of real boiling physics
    The dataset is generated by Flash-X, and validation against experiments is only qualitative (flow regime agreement in Appendix B.7); the paper defers quantitative comparison to future work. The downstream forecasting claims inherit this premise.
  • standard math Two-phase incompressible Navier-Stokes with level-set interface tracking is the correct physics
    Adopted from prior solver formulations [10, 21, 30]; the question of whether the simulated physics matches real boiling is separate, but the governing equations are standard for this regime.
  • domain assumption Interpolation of AMR data to regular grids preserves the relevant boiling dynamics
    Appendix Section 5 states AMR simulations are interpolated to regular grids; the paper does not quantify interpolation errors for the physics of interest.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Bubbleformer: Forecasting Boiling with Transformers." pith.science (2026). https://pith.science/paper/BGTIO6JT

@misc{pith2026250721244,
  author       = {Pith},
  title        = {Pith review of: Bubbleformer: Forecasting Boiling with Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BGTIO6JT}},
  note         = {Machine review of arXiv:2507.21244}
}
read the original abstract

Modeling boiling (an inherently chaotic, multiphase process central to energy and thermal systems) remains a significant challenge for neural PDE surrogates. Existing models require future input (e.g., bubble positions) during inference because they fail to learn nucleation from past states, limiting their ability to autonomously forecast boiling dynamics. They also fail to model flow boiling velocity fields, where sharp interface-momentum coupling demands long-range and directional inductive biases. We introduce Bubbleformer, a transformer-based spatiotemporal model that forecasts stable and long-range boiling dynamics including nucleation, interface evolution, and heat transfer without dependence on simulation data during inference. Bubbleformer integrates factorized axial attention, frequency-aware scaling, and conditions on thermophysical parameters to generalize across fluids, geometries, and operating conditions. To evaluate physical fidelity in chaotic systems, we propose interpretable physics-based metrics that evaluate heat-flux consistency, interface geometry, and mass conservation. We also release BubbleML 2.0, a high-fidelity dataset that spans diverse working fluids (cryogens, refrigerants, dielectrics), boiling configurations (pool and flow boiling), flow regimes (bubbly, slug, annular), and boundary conditions. Bubbleformer sets new benchmark results in both prediction and forecasting of two-phase boiling flows.

Figures

Figures reproduced from arXiv: 2507.21244 by the authors.

Figure 1
Figure 1. (a) BubbleML 2.0 dataset: Visualizes temperature and phase fields across pool and flow boiling configurations. Bubbles nucleate on heated surfaces, deform, coalesce, and transition between regimes, with dynamics strongly coupled to boundary conditions, working fluid, and geometry. (b) Bubbleformer downstream tasks: Autonomous forecasting of full-field dynamics including nucleation and interface evolution, and predic… view at source ↗
Figure 2
Figure 2. Failure modes of current ML models in forecasting. Left: UNet models maintain smooth evolution during autoregressive rollout but fail to nucleate. Right: FNO exhibits early instability. distance field (due to the newly formed bubble) with the evolving phase field, introducing sharp discontinuities and topological changes. These reinitializations are non-differentiable and violate assumptions of spatial smoothness. A… view at source ↗
Figure 3
Figure 3. Bubbleformer architecture. Input fields are encoded into multiscale spatiotemporal patches, conditioned with fluid-specific parameters, transformed by frequency-scaled temporal and spatial axial attention, and decoded into future velocity, temperature, and interface fields. Hierarchical Patch Embedding. Bubbleformer applies a hierarchical MLP (hMLP) stem inspired by recent hybrid vision transformers [49] to embed in… view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Flow Boiling Forecasting. (a) Rollout for a Bubbleformer-S model on an unseen flow boiling trajectory (FC-72, input flow scale = 2.2). (b) Comparison of predicted vs. ground-truth heat flux PDFs. (c) Per-frame Eikonal loss of forecasted signed distance fields represent…
Figure 5
Figure 5. Figure 5: Subcooled Pool Boiling Prediction. (a) and (b) Predicted temperature and velocity fields from the Bubbleformer-S model on an unseen subcooled pool boiling trajectory for FC-72. (c) and (d) Relative L2 error over 800 rollout timesteps for temperature and velocity magnit…
Figure 6
Figure 6. Figure 6: Comparison of flow boiling simulations at constant heat flux with experimental observations. (a) Illustration of a horizontal channel showing different flow boiling regimes [19]. (b) Experimental visualizations of FC-72 at increasing heat flux values expressed as a fra…
Figure 7
Figure 7. Figure 7: Single-Bubble Forecasting for R515B at 33 °C Wall Superheat. Comparison of ground truth and Bubbleformer predicted bubble position (signed distance field), temperature, and velocity fields for an idealized single-bubble scenario. Forecasting was performed using the Bub…
Figure 8
Figure 8. Figure 8: Single-Bubble Forecasting for FC-72 at 33 °C Wall Superheat. Comparison of ground truth and Bubbleformer predicted bubble position (signed distance field), temperature, and velocity fields for an idealized single-bubble scenario. Forecasting was performed using the Bub…
Figure 9
Figure 9. Figure 9: Forecasting Performance: Single Bubble Pool Boiling. Metrics shown across varying wall superheats and timesteps for R515B and FC-72 under saturated pool boiling conditions. We report bubble growth time(ms), heat flux prediction (system-level quantity), Eikonal loss (in…
Figure 10
Figure 10. Figure 10: Forecasting Results: Saturated Pool Boiling. (a) Rollout for a Bubbleformer-L model on an unseen pool boiling trajectory, FC-72 at heater temperature= 91 °C. (b) and (c) Comparison of predicted vs. ground-truth heat flux PDFs for R515B and FC-72 respectively at differ…
Figure 11
Figure 11. Figure 11: Forecasting Results: Subcooled Pool Boiling. (a) Rollout for a Bubbleformer-L model on an unseen pool boiling trajectory, R515B at heater temperature= 30 °C. (b) and (c) Comparison of predicted vs. ground-truth heat flux PDFs for R515B and FC-72 respectively at differ…
Figure 12
Figure 12. Figure 12: Saturated Pool Boiling Prediction. (a) Predicted temperature and velocity fields from the Bubbleformer-S model on an unseen subcooled pool boiling trajectory for R515B. (b) Relative L2 error over 800 rollout timesteps for temperature and velocity magnitude (combined x…
Figure 13
Figure 13. Figure 13: Flow Boiling Inlet Velocity Prediction. (a) Predicted temperature and velocity fields from the Bubbleformer-S model on an unseen flow boiling trajectory for FC-72. (b) Relative L2 error over 800 rollout timesteps for temperature and velocity magnitude (combined x and …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

62 extracted references · 57 canonical work pages

  1. [1]

    Arnab, M

    A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Luˇci´c, and C. Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision , pages 6836–6846, 2021

    work page 2021

  2. [2]

    N. Basu, G. R. Warrier, and V . K. Dhir. Wall heat flux partitioning during subcooled flow boiling: Part 1—model development. Journal of Heat Transfer, 127(2):131–140, 03 2005

    work page 2005

  3. [3]

    Bertasius, H

    G. Bertasius, H. Wang, and L. Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), July 2021. 10

    work page 2021

  4. [4]

    C. M. Bishop and H. Bishop. Deep Learning: Foundations and Concepts. Springer, 2024

    work page 2024

  5. [5]

    Brandstetter, D

    J. Brandstetter, D. E. Worrall, and M. Welling. Message passing neural PDE solvers. In International Conference on Learning Representations, 2022

    work page 2022

  6. [6]

    Chauhan, S

    P. Chauhan, S. E. Choutri, M. Ghattassi, N. Masmoudi, and S. E. Jabari. Neural operators struggle to learn complex pdes in pedestrian mobility: Hughes model case study, 2025

    work page 2025

  7. [7]

    X. Chen, C. Liang, D. Huang, E. Real, K. Wang, Y . Liu, H. Pham, X. Dong, T. Luong, C.-J. Hsieh, Y . Lu, and Q. V . Le. Symbolic discovery of optimization algorithms, 2023

    work page 2023

  8. [8]

    Chi-Yeh and P

    H. Chi-Yeh and P. Griffith. The mechanism of heat transfer in nucleate pool boiling—part i: Bubble initiaton, growth and departure. International Journal of Heat and Mass Transfer, 8(6):887–904, 1965

    work page 1965

Show all 62 references
  1. [9]

    Devahdhanush, I

    V . Devahdhanush, I. Mudawar, H. K. Nahra, R. Balasubramaniam, M. M. Hasan, and J. R. Mackey. Experimental heat transfer results and flow visualization of vertical upflow boiling in earth gravity with subcooled inlet conditions – in preparation for experiments onboard the inte...

    2022

  2. [10]

    A. Dhruv. A vortex damping outflow forcing for multiphase flows with sharp interfacial jumps. Journal of Computational Physics, 511:113122, 2024

    2024

  3. [11]

    Dhruv, E

    A. Dhruv, E. Balaras, A. Riaz, and J. Kim. A formulation for high-fidelity simulations of pool boiling in low gravity. International Journal of Multiphase Flow, 120:103099, 2019

    2019

  4. [12]

    Dhruv, E

    A. Dhruv, E. Balaras, A. Riaz, and J. Kim. An investigation of the gravity effects on pool boiling heat transfer via high-fidelity simulations. International Journal of Heat and Mass Transfer, 180:121826, 2021

    2021

  5. [13]

    Dhruv and S

    A. Dhruv and S. M. S. Hassan. Github:Lab-Notebooks/Flash-X-Development: 2025.04

    2025

  6. [14]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021

    2021

  7. [15]

    Dubey, K

    A. Dubey, K. Weide, J. O’Neal, A. Dhruv, S. Couch, J. A. Harris, T. Klosterman, R. Jain, J. Rudi, B. Messer, M. Pajkos, J. Carlson, R. Chu, M. Wahib, S. Chawdhary, P. M. Ricker, D. Lee, K. Antypas, K. M. Riley, C. Daley, M. Ganapathy, F. X. Timmes, D. M. Townsley, M. Vanella, ...

    2022

  8. [16]

    V . Ganesan. Development of a finite volume general two-phase navier-stokes solver for direct numerical simulations on cut-cells with sharp fixed interface. Master’s thesis, Purdue University, 2018

    2018

  9. [17]

    Ganesan, R

    V . Ganesan, R. Patel, J. Hartwig, and I. Mudawar. Review of databases and correlations for saturated flow boiling heat transfer coefficient for cryogens in uniformly heated tubes, and development of new consolidated database and universal correlations. International Journal o...

    2021

  10. [18]

    Ganesan, R

    V . Ganesan, R. Patel, J. Hartwig, and I. Mudawar. Universal critical heat flux (chf) correlations for cryogenic flow boiling in uniformly heated tubes. International Journal of Heat and Mass Transfer, 166:120678, 2021

    2021

  11. [19]

    Ganesan, R

    V . Ganesan, R. Patel, J. Hartwig, and I. Mudawar. Universal correlations for post-chf saturated and superheated flow film boiling heat transfer coefficient, minimum heat flux and rewet temperature for cryogenic fluids in uniformly heated tubes. International Journal of Heat a...

    2022

  12. [20]

    Ganesan, R

    V . Ganesan, R. Patel, J. Hartwig, and I. Mudawar. Development of two-phase frictional pressure gradient correlation for saturated cryogenic flow boiling in uniformly heated tubes.International Journal of Heat and Mass Transfer, 220:124901, 2024. 11

    2024

  13. [21]

    Gibou, L

    F. Gibou, L. Chen, D. Nguyen, and S. Banerjee. A level set based sharp interface method for the multiphase incompressible Navier-Stokes equations with phase change. Journal of Computational Physics, 222(2):536–555, 2007

    2007

  14. [22]

    Gilman and E

    L. Gilman and E. Baglietto. A self-consistent, physics-based boiling heat transfer modeling framework for use in computational fluid dynamics. International Journal of Multiphase Flow, 95:35–53, 2017

    2017

  15. [23]

    J. K. Gupta and J. Brandstetter. Towards multi-spatiotemporal-scale generalized PDE modeling. arXiv preprint arXiv:2209.15616, 2022

    2022 arXiv

  16. [24]

    S. M. S. Hassan, A. Feeney, A. Dhruv, J. Kim, Y . Suh, J. Ryu, Y . Won, and A. Chandramowlish- waran. Bubbleml: A multiphase multiphysics dataset and benchmarks for machine learning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Ne...

    2023

  17. [25]

    Hendrycks and K

    D. Hendrycks and K. Gimpel. Gaussian error linear units (gelus), 2023

    2023

  18. [26]

    J. Ho, N. Kalchbrenner, D. Weissenborn, and T. Salimans. Axial attention in multidimensional transformers, 2019

    2019

  19. [27]

    Huang, X

    Z. Huang, X. Wang, L. Huang, C. Huang, Y . Wei, and W. Liu. Ccnet: Criss-cross attention for semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 603–612, 2019

    2019

  20. [28]

    M. J. Inanlu, V . Ganesan, N. V . Upot, C. Wang, Z. Suo, K. Fazle Rabbi, P. Kabirzadeh, A. Bakhshi, W. Fu, T. S. Thukral, et al. Unveiling the fundamentals of flow boiling heat transfer enhancement on structured surfaces. Science Advances, 10(45):eadp8632, 2024

    2024

  21. [29]

    F. P. Incropera, D. P. DeWitt, T. L. Bergman, and A. S. Lavine.Fundamentals of Heat and Mass Transfer. John Wiley & Sons, New York, 6th edition, 2007

    2007

  22. [30]

    M. Kang, R. P. Fedkiw, and X.-D. Liu. A boundary condition capturing method for multiphase incompressible flow. Journal of Scientific Computing, 15(3):323–360, Sep 2000

    2000

  23. [31]

    Khodakarami, V

    S. Khodakarami, V . Oommen, A. Bora, and G. E. Karniadakis. Mitigating spectral bias in neural operators via high-frequency scaling for physical systems, 2025

    2025

  24. [32]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, and A. Anandkumar. Neural operator: Learning maps between function spaces with applications to pdes. Journal of Machine Learning Research, 24(89):1–97, 2023

    2023

  25. [33]

    N. B. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. M. Stuart, and A. Anand- kumar. Neural operator: Learning maps between function spaces. CoRR, abs/2108.08481, 2021

    2021 arXiv

  26. [34]

    Lanthaler, R

    S. Lanthaler, R. Molinaro, P. Hadorn, and S. Mishra. Nonlinear reconstruction for operator learning of PDEs with discontinuities. In The Eleventh International Conference on Learning Representations, 2023

    2023

  27. [35]

    E. W. Lemmon, I. H. Bell, M. Huber, and M. McLinden. Nist standard reference database 23: reference fluid thermodynamic and transport properties-refprop, version 10.0, national institute of standards and technology. Standard Reference Data Program, Gaithersburg, pages 45–46, 2018

  28. [36]

    Lhoest, A

    Q. Lhoest, A. Villanova del Moral, Y . Jernite, A. Thakur, P. von Platen, S. Patil, J. Chaumond, M. Drame, J. Plu, L. Tunstall, J. Davison, M. Šaško, G. Chhablani, B. Malik, S. Brandeis, T. Le Scao, V . Sanh, C. Xu, N. Patry, A. McMillan-Major, P. Schmid, S. Gugger, C. Delangu...

    2021

  29. [37]

    Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. liu, K. Bhattacharya, A. Stuart, and A. Anand- kumar. Fourier neural operator for parametric partial differential equations. In International Conference on Learning Representations, 2021

    2021

  30. [38]

    Liang and I

    G. Liang and I. Mudawar. Pool boiling critical heat flux (chf)–part 1: Review of mechanisms, models, and correlations. International Journal of Heat and Mass Transfer, 117:1352–1367, 2018

    2018

  31. [39]

    Lippe, B

    P. Lippe, B. S. Veeling, P. Perdikaris, R. E. Turner, and J. Brandstetter. Modeling accurate long rollouts with temporal neural PDE solvers. In ICML Workshop on New Frontiers in Learning, Control, and Dynamical Systems, 2023

    2023

  32. [40]

    McCabe, B

    M. McCabe, B. R.-S. Blancard, L. H. Parker, R. Ohana, M. Cranmer, A. Bietti, M. Eickenberg, S. Golkar, G. Krawezik, F. Lanusse, M. Pettee, T. Tesileanu, K. Cho, and S. Ho. Multiple physics pretraining for spatiotemporal surrogate models. In The Thirty-eighth Annual Conference ...

    2024

  33. [41]

    McCabe, P

    M. McCabe, P. Harrington, S. Subramanian, and J. Brown. Towards stability of autoregressive neural operators. arXiv preprint arXiv:2306.10619, 2023

    2023 arXiv

  34. [42]

    Mukherjee and S

    A. Mukherjee and S. G. Kandlikar. Numerical study of single bubbles with dynamic contact angle during nucleate pool boiling. International Journal of Heat and Mass Transfer, 50(1- 2):127–138, 2007

    2007

  35. [43]

    Perez, F

    E. Perez, F. Strub, H. de Vries, V . Dumoulin, and A. C. Courville. Film: Visual reasoning with a general conditioning layer. In AAAI, 2018

    2018

  36. [44]

    Raffel, N

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y . Zhou, W. Li, and P. J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer, 2023

    2023

  37. [45]

    S. M. Rassoulinejad-Mousavi, F. Al-Hindawi, T. Soori, A. Rokoni, H. Yoon, H. Hu, T. Wu, and Y . Sun. Deep learning strategies for critical heat flux detection in pool boiling.Applied Thermal Engineering, 190:116849, 2021

    2021

  38. [46]

    Sato and B

    Y . Sato and B. Niceno. Nucleate pool boiling simulations using the interface tracking method: Boiling regime from discrete bubble to vapor mushroom region. International Journal of Heat and Mass Transfer, 105:505 – 524, 2017

    2017

  39. [47]

    Sato and B

    Y . Sato and B. Niceno. Pool boiling simulation using an interface tracking method: From nucleate boiling to film boiling regime through critical heat flux. International Journal of Heat and Mass Transfer, 125:876 – 890, 2018

    2018

  40. [48]

    K. N. R. Sinha, V . Kumar, N. Kumar, A. Thakur, and R. Raj. Deep learning the sound of boiling for advance prediction of boiling crisis. Cell Reports Physical Science, 2(3), 2021

    2021

  41. [49]

    Touvron, M

    H. Touvron, M. Cord, A. El-Nouby, J. Verbeek, and H. Jégou. Three things everyone should know about vision transformers, 2022

    2022

  42. [50]

    A. Tran, A. Mathews, L. Xie, and C. S. Ong. Factorized fourier neural operators. In The Eleventh International Conference on Learning Representations, 2023

    2023

  43. [51]

    H. Wang, Y . Zhu, B. Green, H. Adam, A. Yuille, and L.-C. Chen. Axial-deeplab: Stand-alone axial-attention for panoptic segmentation. In European conference on computer vision, pages 108–126. Springer, 2020

    2020

  44. [52]

    P. Wang, W. Zheng, T. Chen, and Z. Wang. Anti-oversmoothing in deep vision transformers via the fourier domain analysis: From theory to practice. In International Conference on Learning Representations, 2022

    2022

  45. [53]

    M. D. Wilkinson, M. Dumontier, I. J. Aalbersberg, G. Appleton, M. Axton, A. Baak, N. Blomberg, J.-W. Boiten, L. B. da Silva Santos, P. E. Bourne, et al. The fair guiding principles for scientific data management and stewardship. Scientific data, 3(1):1–9, 2016

    2016

  46. [54]

    R. J. Williams and D. Zipser. A learning algorithm for continually running fully recurrent neural networks. Neural Computation, 1(2):270–280, 1989. 13

    1989

  47. [55]

    L. Xiao, Y . Zhuang, X. Wu, J. Yang, Y . Lu, Y . Liu, and X. Han. A review of pool-boiling processes based on bubble-dynamics parameters. Applied Sciences, 13(21), 2023

    2023

  48. [56]

    H. Yang, Y . Sun, G. Sundaramoorthi, and A. Yezzi. Steik: Stabilizing the optimization of neural signed distance functions and finer shape representation. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing...

    2023

  49. [57]

    Yazdani, T

    M. Yazdani, T. Radcliff, M. Soteriou, and A. A. Alahyari. A high-fidelity approach towards simulation of pool boiling. Physics of Fluids, 28(1):1–30, 2016

    2016

  50. [58]

    L. Zeng, J. Klausner, and R. Mei. A unified model for the prediction of bubble detachment diameters in boiling systems—i. pool boiling. International Journal of Heat and Mass Transfer, 36(9):2261–2270, 1993

    1993

  51. [59]

    Zhang, M

    P. Zhang, M. P. Laiu, M. Norman, D. Stefanski, and J. Gounley. Matey: multiscale adaptive foundation models for spatiotemporal physical systems, 2024

    2024

  52. [60]

    X. Zhao, K. Shirvan, R. K. Salko, and F. Guo. On the prediction of critical heat flux us- ing a physics-informed machine learning-aided framework. Applied Thermal Engineering , 164:114540, 2020. 14 A Numerical Formulation Boiling occurs when a liquid undergoes evaporation on t...

    2020

  53. [61]

    UNetmod: UNet is a commonly used image-to-image architecture in computer vision tasks such as image segmentation. While standard UNet models are not specifically designed for PDE learning—especially when training data come from numerical simulations with varying spatial resolu...

  54. [62]

    Neural operators aim to approximate solution operators of PDEs

    F-FNO: F-FNO (Factorized Fourier Neural Operator) [ 50] is a type of neural operator designed to efficiently solve PDEs by learning mappings between function spaces. Neural operators aim to approximate solution operators of PDEs. Given an initial condition u0, a neural operato...

    2000

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.