Pith. sign in

REVIEW 4 major objections 5 minor 60 references

Minimizing a finite-volume residual on the mesh is enough to train an attention GNN for 3D thermo-fluid prediction without labeled CFD data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 10:09 UTC pith:M2NA7LSM

load-bearing objection A serious, well-controlled demonstration that FVM-residual training can replace labeled data for 3D thermo-fluid surrogates; main open question is residual fidelity to the reference solver. the 4 major comments →

arxiv 2607.20321 v2 pith:M2NA7LSM submitted 2026-07-22 physics.flu-dyn cs.LG

Label-Free Finite-Volume-Residual Training of Attention Graph Neural Networks for Coupled Thermo-Fluid Fields

Tianyu Li , Zhiwei Cao , Qingang Zhang , Ruihang Wang , Binyang Song , Yonggang Wen This is my paper
classification physics.flu-dyn cs.LG
keywords finite volume methodresidual-based trainingattention graph neural networklabel-free surrogatethermo-fluid simulationpressure-velocity decouplingunstructured mesh3D CFD surrogate
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper seeks to prove that a neural surrogate for coupled three-dimensional flow and temperature fields can be trained without any precomputed CFD solutions, using only the residual of the finite-volume discretization evaluated on the mesh as the loss. The authors argue that this is possible because the discrete residual is a label-free signal that already encodes the physics, and they back it up with benchmarks on a lid-driven cavity, natural convection, a heated cylinder wake, and a synthetic data center hall. If correct, the approach removes the data-generation bottleneck that dominates the cost of building neural surrogates, while matching or beating a supervised baseline on held-out conditions. The paper's core claim is that the finite-volume residual alone is a sufficient training signal for coupled thermo-fluid prediction.

Core claim

The central discovery is that minimizing a GPU-accelerated differentiable finite-volume residual on the mesh itself is a sufficient training objective for an attention graph neural network to learn coupled velocity, pressure, and temperature fields. The loss is not a continuous PDE residual evaluated through automatic differentiation; it is the actual discrete algebraic balance a finite-volume solver enforces—cell-wise conservation of mass, momentum, and energy—assembled with precomputed sparse mesh operators. With this objective, the model achieves all-field nRMSE of 2.3–2.8% on two steady benchmarks, reproduces buoyancy–energy coupling across Rayleigh numbers, and on two parametric unstead

What carries the argument

The load-bearing object is the finite-volume residual loss LFVM, which evaluates the discrete mass, momentum, and energy balances of Eqs. (7)–(9) on the mesh via precomputed weighted-least-squares gradient reconstruction, face-value interpolation, and conservative flux assembly. Because gradient-based optimization explores directions a residual barely sees, the loss appends a second, face-normal momentum residual (Eq. 15) to the cell momentum residual; the two are nearly blind to complementary grid-parity modes—alternating pressure and velocity fields—and only their combination suppresses pressure–velocity decoupling (Section 2.6). The compound-cell graph, which unifies interior control volu

Load-bearing premise

The claim assumes that the residual as implemented is faithful enough to the true discrete equations that a low residual guarantees a physically correct solution—even though the implementation omits some solver details like limiters, tolerances, and under-relaxation.

What would settle it

Construct a state on the same mesh that has near-zero FVM residual under the combined loss but differs measurably from the CFD reference (e.g., a manufactured non-physical mode orthogonal to both residual families, such as a simultaneous out-of-phase parity pressure and velocity field). If gradient-descent training converges to such a state, the sufficiency of the FVM-residual objective is refuted; observing that all low-residual attractors are accurate would instead support it.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, neural surrogates for coupled 3D thermo-fluid problems can be trained with zero labeled CFD data, removing the dominant upfront data-generation cost and its storage footprint.
  • The residual-trained model matches the CFD reference on steady benchmarks (2.3–2.8% all-field nRMSE) and beats the supervised baseline on held-out parametric conditions for the unsteady cases (rollout nRMSE 0.056 vs 0.076 and 0.119 vs 0.195).
  • The combined cell-plus-face momentum residual provides a systematic way to suppress grid-parity artifacts (checkerboard pressure, velocity stripes) that otherwise appear during residual-based training; the two residuals penalize complementary mode families.
  • The approach is best suited to repeated-simulation workflows: the amortization analysis shows the label-free model reaches cost break-even with CFD sooner than the supervised baseline, though it takes longer to train per epoch.
  • The same residual objective doubles as a per-cell diagnostic of conservation error during training, giving direct feedback on which regions of the domain are least physical.

Where Pith is reading between the lines

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

  • Because the residual cost scales with mesh rather than network size, the same label-free recipe could extend to other PDE families with computable discrete residuals (e.g., structural mechanics, electromagnetics), provided one first checks for near-flat modes with the visibility analysis.
  • The observed blind-mode complementarity suggests a general design rule for physics-residual training: before trusting a residual on a new mesh or equation set, measure which structured modes each residual component penalizes and add components until every tested mode is visible.
  • The gap between residual-trained and supervised baselines on the data center hall hints that conservation coupling acts as a regularizer; comparing the same GNN trained with a supervised loss plus a residual penalty would isolate whether the benefit comes from the objective or from the physics.
  • If the approach extends to wall-modeled or transient boundary conditions—the limitations the authors explicitly name—it would make label-free surrogates practical for design exploration in data centers and similar applications.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a label-free training method for an attention graph neural network that acts as a surrogate for coupled 3D thermo-fluid fields. Instead of regression against precomputed CFD labels, the network is trained by minimizing a differentiable finite-volume residual of the incompressible Navier-Stokes and energy equations, evaluated directly on the mesh. The method is exercised on four benchmarks: a 3D lid-driven cavity, natural convection in a square cavity, an unsteady heated cylinder wake, and a synthetic data center hall. On the two steady benchmarks the reported all-field nRMSE is 2.3-2.8%; on the two transient parametric benchmarks the FVM-loss model is reported to outperform a same-architecture supervised baseline in rollout error (e.g. 0.056 vs 0.076 and 0.119 vs 0.195), with zero data-generation cost. The methods section defines the compound-cell graph, WLSQ gradient reconstruction, face flux assembly, a combined cell/face momentum residual, the log-weighted residual objective, and the data-pool reset schedule.

Significance. If the central claim holds, this is a valuable contribution to scientific machine learning: it would show that a discrete finite-volume residual can supply a sufficient training signal for a neural surrogate of 3D coupled thermo-fluid flows without labeled data, while retaining the conservative structure of the FVM. The paper has several genuine strengths: the controlled comparison against a supervised baseline with identical architecture and training cycle is well designed; the decoupling analysis in Section 2.6 is a careful and informative diagnostic; the data-pool schedule addresses an important rollout-distribution issue; and the authors state code and data availability. The main uncertainties lie in the fidelity of the differentiable residual to the reference solver and in the sensitivity of the results to empirically chosen loss weights. These need to be addressed before the sufficiency claim can be accepted without qualification.

major comments (4)
  1. [Section 4.6 / Eqs. (7)-(11)] The central link between the training objective and the reference solver is asserted but not directly verified. Section 4.6 states that the differentiable residual is not constrained to reproduce every numerical scheme choice, including limiters, linear-solver tolerance, under-relaxation, and PISO/PIMPLE outer-loop settings. Yet the paper never reports the value of the differentiable residual evaluated on the CFD reference states themselves. If converged OpenFOAM states have a large residual under the differentiable operator, then minimizing that operator is solving a different discrete problem, and the reported agreement with CFD would be evidence of architectural inductive bias rather than of the residual objective. I request a direct diagnostic: evaluate the components of Eq. (10)/Eq. (11) on the CFD reference states for held-out test conditions and compare them with the residuals att
  2. [Section 4.5.1 / Eq. (11)] The loss weights in Eq. (11) are empirical scale factors (wcont=10^5, wmom=wenergy=10^4, wpRef=1), and the convective time-weighting theta in Eq. (16) is a default choice. No sensitivity study is provided. Because the training signal is a weighted sum of the conservation residuals, different weights (or a different theta) change the Pareto balance among mass, momentum, and energy residuals and could change the trained solution. The label-free claim does not imply parameter-free, but the robustness of the method to these choices needs to be shown. Please report a sensitivity study around the defaults, for example perturbing the weights by factors of 0.1 and 10 and varying theta in {0, 1/2, 1} on at least one benchmark, with rollout nRMSE as the metric. Without this, it is difficult to exclude the possibility that the reported accuracy depends on a tuned objective rather than on the residu
  3. [Section 2.6 / Figure 22] The decoupling analysis convincingly demonstrates that the cell and face momentum residuals are complementary for the tested grid-parity pressure and velocity modes, and the single-residual ablations are useful. However, the manuscript's language suggests a stronger conclusion: that the combined objective leaves no nearly blind direction. The analysis tests two mode families on two mesh configurations; it does not rule out other unphysical modes that could be nearly invisible to the combined residual. I recommend either (i) adding a broader spectral or adversarial-mode analysis, e.g. injecting randomized or frequency-swept perturbations and measuring the response of the combined residual for each channel, or (ii) softening the claim in Sections 2.6 and 3 to state that the two known parity families are covered, rather than implying general robustness.
  4. [Section 2.7 / Tables 4-5] The amortization comparison is presented in wall-clock time, but the resource basis is not commensurable. The FVM-loss model uses eight RTX 4090 GPUs for roughly 200 hours, whereas the CFD data-generation times are reported as single 4-core CPU solve sums. If data generation is parallelizable over many CPU cores, the wall-clock comparison depends on how many cores are assumed available; if GPUs are more costly per hour than CPUs, the monetary cost comparison differs. To support the claim of lower total model-development cost, the authors should state the hardware assumptions (CPU core count, GPU count, parallel efficiency) and, ideally, report normalized compute cost or energy. At a minimum, the current table should be caveated as wall-clock under the stated hardware configuration, not as a general cost claim.
minor comments (5)
  1. [Section 4.4] The subsection title contains a typo: 'F ace-value reconstruction' should read 'Face-value reconstruction'.
  2. [Section 4.4.4 / Eq. (16)] The notation theta for the convective time-weighting is introduced as distinct from network parameters theta, but the Greek letter is the same in text. Consider using a different symbol (e.g. alpha or eta) to avoid confusion in equations and in oral presentations.
  3. [Figure 22] In panel (b), the labels 'Hex 32^3 3-axis' and 'Hex 32^3 1-axis' are not self-explanatory. Define these in the caption, and state whether the injection is a pure mode or a superposition.
  4. [Tables 2 and 3] The phrase 'cumulative geometric mean' is unfamiliar. It is defined in the table footnote, but a short formal definition near Eq. (1) in Section 2.1 would improve readability, especially since the later tables rely on it.
  5. [Section 3] The third finding states that the FVM-loss model achieves the earliest break-even among the surrogates considered. This is true only under the wall-clock assumptions in Tables 4-5; see the major comment on resource comparability. The text should hedge this claim accordingly.

Circularity Check

0 steps flagged

No load-bearing circularity; FVM residual is defined from governing equations and validated against external CFD references.

full rationale

The claimed derivation—training an attention GNN by minimizing the finite-volume residual (Eqs. 7–11)—does not reduce to its inputs. The residual is assembled from the governing equations (Eqs. 2–4) via WLSQ gradients (Eq. 13), face interpolation (Eq. 14), and conservative flux assembly; no term in L_FVM is fit to the CFD reference fields used for evaluation. The hand-chosen weights (w_cont, w_mom, w_energy, w_pRef) and pressure-gauge term (Eq. 12) are tuning constants, not fitted parameters that later masquerade as predictions. Evaluation against CFD and de Vahl Davis [37] is external benchmark comparison, not training supervision. Self-citations [7] and [25] are contextual prior-work references and are not load-bearing: the central premise rests on the paper's own residual construction and the reported external comparisons. The main admitted weakness—Section 4.6's statement that the differentiable residual is 'not constrained to reproduce every numerical scheme choice (e.g. specific limiter, linear-solver tolerance, under-relaxation, or PIMPLE/PISO outer-loop settings) down to the last numerical detail'—is an unverified solver-fidelity/correctness gap (e.g., whether reference OpenFOAM states have near-zero differentiable residual), not a circularity: even a biased residual objective is defined independently of the validation targets. Thus no specific reduction of the claimed result to its inputs can be exhibited.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The central claim rests on the residual objective being a faithful discrete representation of the physics and on the optimizer not exploiting near-null directions of that objective. The paper provides strong empirical evidence and a partial operator-level analysis, but no formal guarantee. No new physical entities are introduced; the compound cells and slice-attention tokens are computational constructs, not physical postulates.

free parameters (4)
  • FVM-loss weighting coefficients w = w_cont=1e5, w_mom=1e4, w_energy=1e4, w_pRef=1
    Eq. (11) uses fixed empirical scale factors to balance residual magnitudes; chosen by hand and no sensitivity analysis is reported.
  • Convective time-weighting theta = 0.5 default for cylinder and data center; 1.0 fully implicit option
    Eq. (16) blends old/new-time convective fluxes; theta = 1/2 is a hand-chosen Crank-Nicolson-type scheme that affects training stability and dynamics.
  • Pressure gauge reference p_ref_rgh = not specified in paper
    Eq. (12) constrains the volume-averaged modified pressure to a reference value; if this value is chosen per benchmark to match the CFD gauge, it could make the pressure comparison artificially favorable.
  • Data-pool size and reset rate = 2400 slots; 5 case resets per epoch
    Section 4.5.1: pool size and reset cadence are hand-chosen to produce near-uniform occupancy over the 500-step horizon; the training distribution shapes generalization.
axioms (4)
  • domain assumption The governing physics are correctly modeled by 3D incompressible Navier-Stokes with Boussinesq buoyancy and energy transport (Eqs. 2-4).
    All four benchmarks are laminar, single-phase, constant-property flows. Turbulence, compressibility, and multiphase effects are outside the stated scope.
  • domain assumption The differentiable FVM residual (Eqs. 7-11, 13-15) is a sufficiently faithful discrete approximation of the reference OpenFOAM discretization.
    Section 4.6 explicitly states that limiter, linear-solver tolerance, under-relaxation, and PIMPLE/PISO loop settings are not reproduced. If the residual deviates materially, the label-free signal targets a different discrete problem.
  • ad hoc to paper Minimizing the log-weighted residual norm (Eq. 11) drives the network toward physical discrete solutions, with no important unpenalized modes.
    Section 2.6 demonstrates that parity modes require the additional face residual; there is no proof that other near-null directions of the combined residual are absent.
  • ad hoc to paper The anti-overfitting reset schedule (Section 4.5.1) produces a training distribution over rollout positions and conditions representative of inference.
    Uniform occupancy of the time-step horizon and mixing of training conditions is engineered by the pool/reset design; its effect is validated only empirically on the two parametric benchmarks.

pith-pipeline@v1.3.0-alltime-deepseek · 33401 in / 16126 out tokens · 142962 ms · 2026-08-01T10:09:11.909769+00:00 · methodology

0 comments
read the original abstract

Neural surrogates are widely used in scientific machine learning for fast prediction of three-dimensional (3D) thermo-fluid fields. However, generating training data using conventional numerical solvers often incurs substantial computational and storage costs. We propose to train an attention graph neural network by minimizing the finite-volume method (FVM) residuals of the governing equations. These residuals are evaluated directly on the mesh, requiring no labeled data. We evaluate the trained surrogates against computational fluid dynamics (CFD) references and a data-supervised baseline across four scenarios. On the two steady-state benchmarks, the FVM-loss model achieves an all-field normalized root-mean-square error (nRMSE) of 2.3-2.8%. It demonstrates close agreement with the CFD references, including the buoyancy-energy coupling. On the two parametric transient cases, the FVM-loss model outperforms the supervised baseline in terms of accuracy, while avoiding the data-generation cost entirely. These results indicate that the FVM loss can provide a practical training signal for neural surrogates and reduce the model development cost.

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

60 extracted references · 16 linked inside Pith

  1. [1]

    & Wong, H

    Koomey, J., Berard, S., Sanchez, M. & Wong, H. Implications of historical trends in the electrical efficiency of computing.IEEE Annals of the History of Computing 33, 46–54 (2011)

    2011

  2. [2]

    & Koomey, J

    Masanet, E., Shehabi, A., Lei, N., Smith, S. & Koomey, J. Recalibrating global data center energy-use estimates.Science367, 984–986 (2020)

    2020

  3. [3]

    R., Cruz, E

    Schmidt, R. R., Cruz, E. E. & Iyengar, M. K. Challenges of data center thermal management.IBM Journal of Research and Development49, 709–723 (2005)

    2005

  4. [4]

    Latif, I.et al.Single-node power demand during AI training: Measurements on an 8-GPU NVIDIA H100 system.IEEE Access13, 61740–61747 (2025)

    2025

  5. [5]

    Wang, R.et al.Toward data center digital twins via knowledge-based model cali- bration and reduction.ACM Transactions on Modeling and Computer Simulation (2023)

    2023

  6. [6]

    URL https://arxiv.org/abs/2511.11722

    Sarkar, S., Guillen-Perez, A., Carmichael, Z., Naug, A.et al.Fast 3D surrogate modeling for data center thermal management.arXiv preprint arXiv:2511.11722 (2025). URL https://arxiv.org/abs/2511.11722

    arXiv 2025

  7. [7]

    & Wen, Y

    Cao, Z., Wang, R., Zhou, X. & Wen, Y. Reducio: Model reduction for data center predictive digital twins via physics-guided machine learning.Proceedings of the 9th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (BuildSys ’22)1–10 (2022). 44

    2022

  8. [8]

    International Conference on Learning Representations (ICLR)(2021)

    Li, Z.et al.Fourier neural operator for parametric partial differential equations. International Conference on Learning Representations (ICLR)(2021). URL https://arxiv.org/abs/2010.08895

    Pith/arXiv arXiv 2021

  9. [9]

    URL https://arxiv

    Li, Z.et al.Geometry-informed neural operator for large-scale 3d PDEs.Advances in Neural Information Processing Systems (NeurIPS)(2023). URL https://arxiv. org/abs/2309.00583

    Pith/arXiv arXiv 2023

  10. [10]

    & Karniadakis, G

    Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning nonlinear oper- ators via DeepONet based on the universal approximation theorem of operators. Nature Machine Intelligence3, 218–229 (2021)

    2021

  11. [11]

    & Iorio, F

    Guo, X., Li, W. & Iorio, F. Convolutional neural networks for steady flow approx- imation.Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining481–490 (2016)

    2016

  12. [12]

    & Kempf, A

    Leer, M. & Kempf, A. Fast flow field estimation for various applications with a universally applicable machine learning concept.Flow, Turbulence and Combustion107, 175–200 (2021)

    2021

  13. [13]

    & Xiao, H

    Duraisamy, K., Iaccarino, G. & Xiao, H. Turbulence modeling in the age of data. Annual Review of Fluid Mechanics51, 357–377 (2019)

    2019

  14. [14]

    URL https://arxiv.org/abs/2002.09405

    Sanchez-Gonzalez, A.et al.Learning to simulate complex physics with graph networks.International Conference on Machine Learning (ICML)8459–8468 (2020). URL https://arxiv.org/abs/2002.09405

    Pith/arXiv arXiv 2020

  15. [15]

    & Karniadakis, G

    Raissi, M., Perdikaris, P. & Karniadakis, G. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational Physics378, 686–707 (2019)

    2019

  16. [16]

    & Karniadakis, G

    Raissi, M., Yazdani, A. & Karniadakis, G. E. Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations.Science367, 1026–1030 (2020)

    2020

  17. [17]

    & Ricken, T

    Mandl, L., Goswami, S., Lambers, L. & Ricken, T. Separable physics-informed deeponet: Breaking the curse of dimensionality in physics-informed machine learn- ing.arXiv preprint arXiv:2407.15887(2024). URL https://arxiv.org/abs/2407. 15887

    Pith/arXiv arXiv 2024

  18. [18]

    & Hong, L

    Li, G., Ran, S., Yang, W. & Hong, L. Improving generalization ability of deep-learning-based ODE solvers using continuous dependence.npj Artificial Intelligence1, 22 (2025)

    2025

  19. [19]

    E.et al.Physics-informed machine learning.Nature Reviews Physics3, 422–440 (2021)

    Karniadakis, G. E.et al.Physics-informed machine learning.Nature Reviews Physics3, 422–440 (2021). 45

    2021

  20. [20]

    & Perdikaris, P

    Wang, S., Yu, X. & Perdikaris, P. When and why pinns fail to train: A neu- ral tangent kernel perspective.Journal of Computational Physics449, 110768 (2022)

    2022

  21. [21]

    & Zhang, W

    Cao, W. & Zhang, W. An analysis and solution of ill-conditioning in physics- informed neural networks.Journal of Computational Physics520, 113494 (2025)

    2025

  22. [22]

    D., Chen, D

    Toscano, J. D., Chen, D. T., Oommen, V., Darbon, J. & Karniadakis, G. E. A variational framework for residual-based adaptivity in neural PDE solvers and operator learning.npj Artificial Intelligence2, 32 (2026)

    2026

  23. [23]

    & Darwish, M.The Finite Volume Method in Compu- tational Fluid Dynamics: An Advanced Introduction with OpenFOAM and Matlab (Springer, 2016)

    Moukalled, F., Mangani, L. & Darwish, M.The Finite Volume Method in Compu- tational Fluid Dynamics: An Advanced Introduction with OpenFOAM and Matlab (Springer, 2016)

    2016

  24. [24]

    Versteeg, H. K. & Malalasekera, W.An introduction to computational fluid dynamics: the finite volume method2nd edn (Pearson Education, 2007)

    2007

  25. [25]

    Li, T.et al.Learning to solve pdes with finite volume-informed neural networks in a data-free approach.Journal of Computational Physics530, 113919 (2025)

    2025

  26. [26]

    & Battaglia, P

    Pfaff, T., Fortunato, M., Sanchez-Gonzalez, A. & Battaglia, P. W. Learning mesh- based simulation with graph networks.International Conference on Learning Representations (ICLR)(2021). URL https://arxiv.org/abs/2010.03409

    Pith/arXiv arXiv 2021

  27. [27]

    Brandstetter, J., Worrall, D. E. & Welling, M. Message passing neural PDE solvers.International Conference on Learning Representations (ICLR)(2022). URL https://arxiv.org/abs/2202.03376

    Pith/arXiv arXiv 2022

  28. [28]

    L., Noack, B

    Brunton, S. L., Noack, B. R. & Koumoutsakos, P. Machine learning for fluid mechanics.Annual Review of Fluid Mechanics52, 477–508 (2020)

    2020

  29. [29]

    & Yahav, E

    Alon, U. & Yahav, E. On the bottleneck of graph neural networks and its practi- cal implications.International Conference on Learning Representations (ICLR) (2021). URL https://arxiv.org/abs/2006.05205

    Pith/arXiv arXiv 2021

  30. [30]

    URL https://arxiv

    Praditia, T.et al.Finite volume neural network: Modeling subsurface contam- inant transport.arXiv preprint arXiv:2104.06010(2021). URL https://arxiv. org/abs/2104.06010

    Pith/arXiv arXiv 2021

  31. [31]

    Zhang, H., Xia, X.-L., Wu, Z.-d. & Li, X. Physics-informed graph neural network based on the finite volume method for steady incompressible laminar convective heat transfer.Physics of Fluids37, 013625 (2025)

    2025

  32. [32]

    & Thuerey, N

    Um, K., Brand, R., Fei, Y., Holl, P. & Thuerey, N. Solver-in-the-loop: Learning from differentiable physics to interact with iterative PDE-solvers.Advances in Neural Information Processing Systems (NeurIPS)(2020). URL https://arxiv. 46 org/abs/2007.00016

    Pith/arXiv arXiv 2020

  33. [33]

    Proceedings of the National Academy of Sciences118, e2101784118 (2021)

    Kochkov, D.et al.Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences118, e2101784118 (2021)

    2021

  34. [34]

    & Wang, J.-X

    Du, P., Li, Y., Xu, M. & Wang, J.-X. DiFVM: A vectorized graph-based finite volume solver for differentiable CFD on unstructured meshes.arXiv preprint arXiv:2603.15920(2026). URL https://arxiv.org/abs/2603.15920

    arXiv 2026

  35. [35]

    & Koumoutsakos, P

    Karnakov, P., Litvinov, S. & Koumoutsakos, P. Solving inverse problems in physics by optimizing a discrete loss: Fast and accurate learning without neural networks.PNAS Nexus3, pgae005 (2024)

    2024

  36. [36]

    & Thuerey, N

    K¨ ohler, F., Niedermayr, S., Westermann, R. & Thuerey, N. APEBench: A benchmark for autoregressive neural emulators of PDEs.Advances in Neural Information Processing Systems (NeurIPS)(2024). URL https://arxiv.org/abs/ 2411.00180

    Pith/arXiv arXiv 2024

  37. [37]

    Natural convection of air in a square cavity: a bench mark numerical solution.International Journal for Numerical Methods in Fluids3, 249–264 (1983)

    de Vahl Davis, G. Natural convection of air in a square cavity: a bench mark numerical solution.International Journal for Numerical Methods in Fluids3, 249–264 (1983)

    1983

  38. [38]

    S., Perdikaris, P., Turner, R

    Lippe, P., Veeling, B. S., Perdikaris, P., Turner, R. E. & Brandstetter, J. PDE- refiner: Achieving accurate long rollouts with neural PDE solvers.Advances in Neural Information Processing Systems (NeurIPS)(2023). URL https://arxiv. org/abs/2308.05732

    Pith/arXiv arXiv 2023

  39. [39]

    Williamson, C. H. K. Vortex dynamics in the cylinder wake.Annual Review of Fluid Mechanics28, 477–539 (1996)

    1996

  40. [40]

    & Tao, W.-Q

    Bai, F., Gong, X.-M., Li, H.-W., Guo, H.-B. & Tao, W.-Q. An improved black box model and the details of its numerical treatments for rack in data center sim- ulation.International Communications in Heat and Mass Transfer158, 107916 (2024)

    2024

  41. [41]

    & Zhang, W

    Cao, W. & Zhang, W. Overcoming the loss conditioning bottleneck in optimization-based PDE solvers: a well-conditioned loss function.Communica- tions in Nonlinear Science and Numerical Simulation160, 109952 (2026)

    2026

  42. [42]

    Gray, D. D. & Giorgini, A. The validity of the Boussinesq approximation for liquids and gases.International Journal of Heat and Mass Transfer19, 545–551 (1976)

    1976

  43. [43]

    Wu, Z.et al.A comprehensive survey on graph neural networks.IEEE Transactions on Neural Networks and Learning Systems32, 4–24 (2020). 47

    2020

  44. [44]

    Tancik, M.et al.Fourier features let networks learn high frequency functions in low dimensional domains.Advances in Neural Information Processing Systems (NeurIPS)33, 7537–7547 (2020)

    2020

  45. [45]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O. & Dahl, G. E. Neural message passing for quantum chemistry.International Conference on Machine Learning (ICML)1263–1272 (2017). URL https://arxiv.org/abs/1704.01212

    Pith/arXiv arXiv 2017

  46. [46]

    W.et al.Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261(2018)

    Battaglia, P. W.et al.Relational inductive biases, deep learning, and graph networks.arXiv preprint arXiv:1806.01261(2018). URL https://arxiv.org/abs/ 1806.01261

    Pith/arXiv arXiv 2018

  47. [47]

    & Long, M

    Wu, H., Luo, H., Wang, H., Wang, J. & Long, M. Transolver: A fast transformer solver for PDEs on general geometries.Proceedings of the 41st International Conference on Machine Learning (ICML), PMLR235, 53681–53705 (2024). URL https://arxiv.org/abs/2402.02366

    Pith/arXiv arXiv 2024

  48. [48]

    URL https://arxiv.org/ abs/1706.03762

    Vaswani, A.et al.Attention is all you need.Advances in Neural Information Processing Systems (NeurIPS)30, 5998–6008 (2017). URL https://arxiv.org/ abs/1706.03762

    Pith/arXiv arXiv 2017

  49. [49]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. IEEE Conference on Computer Vision and Pattern Recognition (CVPR)770–778 (2016)

    2016

  50. [50]

    & Perdikaris, P

    Wang, S., Teng, Y. & Perdikaris, P. Understanding and mitigating gradient flow pathologies in physics-informed neural networks.SIAM Journal on Scientific Computing43, A3055–A3081 (2021)

    2021

  51. [51]

    Mavriplis, D. J. Revisiting the least-squares procedure for gradient reconstruction on unstructured meshes.16th AIAA Computational Fluid Dynamics Conference, AIAA Paper 2003-3986(2003)

    2003

  52. [52]

    Harlow, F. H. & Welch, J. E. Numerical calculation of time-dependent viscous incompressible flow of fluid with free surface.The Physics of Fluids8, 2182–2189 (1965)

    1965

  53. [53]

    Rhie, C. M. & Chow, W. L. Numerical study of the turbulent flow past an airfoil with trailing edge separation.AIAA Journal21, 1525–1532 (1983)

    1983

  54. [54]

    Conservation properties of unstructured staggered mesh schemes

    Perot, B. Conservation properties of unstructured staggered mesh schemes. Journal of Computational Physics159, 58–89 (2000)

    2000

  55. [55]

    Min, Q.et al.Efficient neural network training method for unsteady flow field prediction based on data pool.Engineering Applications of Computational Fluid Mechanics19, 2512958 (2025). 48

    2025

  56. [56]

    URL https://arxiv.org/abs/2409.11321

    Vyas, N.et al.SOAP: Improving and stabilizing shampoo using adam for lan- guage modeling.International Conference on Learning Representations (ICLR) (2025). URL https://arxiv.org/abs/2409.11321

    Pith/arXiv arXiv 2025

  57. [57]

    & Thuerey, N

    List, B., Chen, L.-W., Bali, K. & Thuerey, N. Differentiability in unrolled training of neural physics simulators on transient dynamics.Computer Methods in Applied Mechanics and Engineering433, 117441 (2025)

    2025

  58. [58]

    LeCun, Y., Bottou, L., Orr, G. B. & M¨ uller, K.-R. Efficient backprop.Neural Networks: Tricks of the Trade, Lecture Notes in Computer Science1524, 9–50 (1998)

    1998

  59. [59]

    G., Tabor, G., Jasak, H

    Weller, H. G., Tabor, G., Jasak, H. & Fureby, C. A tensorial approach to com- putational continuum mechanics using object-oriented techniques.Computers in Physics12, 620–631 (1998)

    1998

  60. [60]

    Issa, R. I. Solution of the implicitly discretised fluid flow equations by operator- splitting.Journal of Computational Physics62, 40–65 (1986). Appendix A Structure and conditioning of the residual Jacobian After boundary-condition injection on the ghost cells, the discrete residual operator acting on the updated stateq n+1 θ =q n + ∆qθ (Eq. (6)) is rθ =...

    1986