Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Physics-based machine learning for mantle convection simulations

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A learned CNN that predicts divergence-free velocities can replace the Stokes solver in mantle convection, producing stable rollouts and up to 89x speedup.

desk verdict Solid engineering: first CNN Stokes surrogate for mantle convection that rolls out stably, with honest OOD tests; main gap is the unquantified velocity scaler behind magnitude recovery. read the letter →

arxiv 2505.16041 v1 pith:Y5KXAXV4 submitted 2025-05-21 astro-ph.EP cs.LG

Siddhant Agarwal , Ali Can Bekar , Christian Hüttig , David S. Greenberg , Nicola Tosi This is my paper · ORCID
classification astro-ph.EPcs.LG
keywords mantleconvectionStokesflowsurrogateconvolutionalneuralnetworkmassconservationbyconstructionautoregressiverolloutphysics-basedmachinelearningplanetarythermalevolutionincompressiblecreeping
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

This paper tries to show that the most expensive part of a mantle convection simulation—the numerical solution of the Stokes flow equations—can be replaced by a compact convolutional network trained on only 94 simulations. The network predicts divergence-free velocities directly from the temperature field, and a finite-volume solver advects and diffuses temperature with those velocities. The authors report stable autoregressive rollouts over tens of thousands of time steps and a speedup of up to 89x against the direct solver. The claim matters because planetary thermal evolution models require many long simulations with poorly constrained parameters; cutting the cost of the momentum solve makes broad parameter studies more feasible.

What carries the argument

The central object is the curl constraint $u = \partial a/\partial y$, $v = -\partial a/\partial x$, where $a$ is a single scalar field learned by the network; forming velocities this way makes the predicted flow divergence-free by construction. Around it sit three supporting mechanisms: a fitted log-linear velocity scaler, Eq. (4), that normalizes velocities across several orders of magnitude; learned boundary convolutions that reduce edge errors and prevent artificial hot bubbles during advection; and a loss scaled by each example's velocity range with a derivative penalty to suppress oscillations in the curl-derived fields. These components are coupled to a numerical finite-volume solver that advances the temperature field, enabling rollout without ever learning in time.

What would settle it

A direct test is to take a point in the high-Q and high-gamma corner outside the training envelope, compare the Eq. (4) scaler against the true maximum velocity from the numerical solver, and run the hybrid rollout; if the mismatch in velocity magnitudes alone predicts the divergence, the scaler is the culprit, while if the rollout fails even with corrected magnitudes, the cause lies elsewhere.

Watch

Extended reading notes

Core claim

The central claim is that a convolutional network, designed along the lines of a learned linear-system surrogate and enforcing incompressibility by predicting a scalar field whose curl gives the velocity components, can stand in for the Stokes solver and be coupled to a finite-volume advection-diffusion step. Mass conservation is exact in the interior to double machine precision and softly enforced on the boundaries. The paper argues that because the momentum equation has no memory—velocities depend only on the current temperature and viscosity—the network can be trained on snapshots and used autoregressively without learning in time. On unseen parameter combinations it produces stable rollouts for tens of thousands of time steps, with velocity predictions about 14 times more accurate than using the last available solver velocities, and speedups up to 89x. The paper also claims that direct time-stepping with a U-net diverges within 16 steps despite lower single-step velocity loss, which it attributes to the U-net being conditioned on previous velocities, and it acknowledges that extrapolation to high internal heating rate and high temperature-dependent viscosity fails.

Load-bearing premise

The load-bearing premise is that the fitted speed-adjustment formula in Eq. (4), estimated from the maximum velocities of 94 training simulations, continues to give correct velocity magnitudes for parameter combinations it never saw; if that scaling is wrong, predicted flows are systematically too fast or too slow and rollouts fail, which the paper observes in the high-Q and high-gamma region.

Editorial extensions

If this is right

  • A single network trained on 94 simulations can predict velocities across four orders of magnitude and roll out stably for tens of thousands of time steps in 16 of 18 unseen test simulations.
  • The hybrid model reaches up to 89x speedup over the direct CPU solver while matching mean-temperature histories more closely than a solver that skips momentum solves every 100 steps.
  • Mass conservation by construction is claimed to be decisive: adding mass conservation only as a soft loss does not provide the same stability.
  • The model can simulate thermal evolution with a time-decaying heat source even though it was trained on steady-state data, but it fails when extrapolating to high Q and high gamma or to a compressible energy equation.
  • Direct time-stepping with a U-net diverges within 16 steps despite lower single-step velocity loss, supporting the paper's argument that the memory-free structure of the momentum equation is the right inductive bias.

Reading between the lines

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

  • If the velocity-scaling idea generalizes, the same hybrid scheme could drop explicit parameter inputs and instead scale velocities from the viscosity field itself, opening the way to other rheologies and geometries.
  • The 89x speedup is measured for a single run on one GPU/CPU pair; the paper notes that Python-C++ data exchange throttles parallel campaigns, so the practical speedup in production parameter studies will likely be lower unless inference is moved into the solver itself.
  • The same division of labor—learned surrogate for the memory-free momentum equation, numerical solver for energy advection-diffusion—could apply to other creeping-flow PDEs, not just mantle convection.
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 proposes a physics-based machine-learning surrogate for the Stokes solve in 2D mantle-convection simulations. A convolutional network takes temperature, viscosity, grid coordinates, and the parameters Q, β, and γ as inputs and predicts a vector potential whose curl gives divergence-free velocity fields. These velocities are multiplied by a fitted velocity scaler (Eq. 4) and fed to the finite-volume solver GAIA, which performs the advection-diffusion step; this enables autoregressive rollout without learning in time. The authors train on 94 simulations, evaluate on 18 unseen simulations, report stable rollouts for 16 of them, and claim an up-to-89x speedup over a CPU direct solver. They also compare against momentum-skip and GPU-iterative baselines, perform ablations of mass conservation, boundary padding, derivative loss, and loss scaling, and test out-of-distribution scenarios including a decaying heat source and the Extended Boussinesq Approximation.

Significance. If the claims hold, this is a useful contribution to accelerating mantle-convection parameter studies. The model is trained on only 94 simulations, conserves mass by construction via a curl formulation, uses learned boundary padding to reduce boundary artifacts, and the authors explicitly report failure cases in extrapolation regions, which is more informative than typical surrogate papers. The public code and dataset availability are concrete strengths. However, the headline claims of stable long rollouts and large speedup currently rest on qualitative or single-run evidence, and the velocity scaler that fixes absolute velocity magnitudes is unquantified. With the requested aggregate statistics and an explicit characterization of the scaler fit, the contribution would be solidly supported.

major comments (3)
  1. [II B, Eq. (4)] The velocity scaler in Eq. (4) is load-bearing for the central claim: every CNN-predicted velocity is multiplied by this factor before being passed to the advection-diffusion solver, so any error in the scaler translates directly into an error in advective transport rates. The manuscript never reports the quality of the log-linear fit on the 94 training simulations, the distribution of residuals, or the ratio of predicted to true maximum velocity on the 18 test simulations. Because the scaler is a power law in Q, γ, and β, extrapolation error is expected at the high-Q, high-γ corner, which is exactly where Fig. 6 locates the two failed or nonphysical rollouts. The paper should quantify the scaler fit (e.g., R², residual range, test-set predicted-versus-true maximum velocities) and, if necessary, assess whether the observed failures are attributable to the scaler or to the CNN's spatial errors.
  2. [III D, Fig. 6] The stability claim ("stable rollout for tens of thousands of time-steps") is not backed by a definition of stability or by aggregate test-set statistics. Only 3 of 18 test simulations are shown in Fig. 5, and the paper reports that 16 of 18 runs are "stable" without giving a criterion (e.g., bounded mean temperature, bounded maximum temperature error, no divergence) or the distribution of errors over the remaining 16 runs. Since the abstract and conclusion foreground stable long rollouts, the authors should report per-simulation rollout MAEs or a summary statistic over the full test set and state the threshold used to classify a run as unstable or nonphysical.
  3. [III E, Fig. 7] The 89x speedup is reported from what appears to be a single benchmark run on one GPU/CPU configuration with no measurement details. The abstract and conclusion rely on this number, but the text does not specify the hardware (beyond V100), the number of repeated runs, or the wall-clock breakdown among CNN inference, Python/C++ data transfer, and the advection-diffusion solve. The authors do note in Sec. III E that parallel runs throttle the speedup; this caveat should be quantified, and the speedup should be reported with an uncertainty or range, together with the exact benchmark protocol.
minor comments (5)
  1. [II A] After Eq. (3), the text contains "where, where" — the duplicated word should be removed.
  2. [II C 4] After Eq. (10), there is an isolated equation label "(11)" with no corresponding displayed equation; this should be removed or renumbered.
  3. [Table I] The Table I caption says "we compute the MAE on the boundaries ... and plot them," but the results are presented in a table, not a plot; the caption wording should be corrected.
  4. [References] Some reference entries contain formatting artifacts, such as "patrick gallinari" in lowercase and corrupted characters in the Um et al. entry; these should be cleaned.
  5. [Data Availability] The data availability section contains "available herehttps://github.com/..." with a missing space, and the acknowledgments contain "the the German Ministry"; both should be corrected.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is an emulator benchmarked on held-out solver data, not a derivation that reduces to its inputs.

full rationale

The paper's claimed chain is an empirical surrogate: GAIA Stokes snapshots are used to train a CNN that outputs divergence-free velocities via a vector potential, the same GAIA code advects temperature with those velocities, and held-out GAIA simulations serve as the benchmark. No step reduces by construction to its input. The velocity scaler in Eq. (4) is a fitted normalization, not a derivation or an evaluation target; the final predicted velocity is scaler times the CNN output, so the network can compensate for scaler error and the rollout is not a re-statement of the regression. Mass conservation is an imposed constraint, not a predicted consequence. The self-citations (Agarwal et al. 2021, 2024, 2025) provide the dataset, an initial-condition tool, and a data-efficiency comparison, but none is load-bearing for the stability or speedup claims, and none is invoked as an external theorem that forces the method. The extrapolation failures at high Q and gamma are a validation limitation, not a circular step.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the physical modeling assumptions (Boussinesq Stokes flow, specific viscosity law) and on a normalization fit to the training data. No new physical entities are introduced.

free parameters (4)
  • velocity scaler (Eq. 4) = 5 e^{0.18Q} gamma^{0.43} beta^{-0.46}
    Fit through the maximum observed velocities as a function of Q, beta, gamma; used to normalize velocity outputs before un-normalizing at inference. If inaccurate for unseen parameters, the predicted velocity magnitudes are wrong.
  • vector potential multiplier = 10
    Empirically determined factor multiplying the vector potential a to keep outputs bounded (Sec. II C 2).
  • loss scaling clip = clip at 10
    Loss scaling term Eq. (6) clips inverse range to [1,10]; chosen by hand.
  • boundary loss weight = 10
    Boundary loss weighting factor in Eq. (7), described by the authors as 'somewhat arbitrary'.
assumptions (4)
  • domain assumption Boussinesq approximation and incompressible Stokes flow (Eq. 1) describe mantle convection.
    The training data and the advection solver both solve these equations; if the physics is wrong, the surrogate inherits the error.
  • domain assumption Viscosity is given by eta(T,y) = exp(-log(gamma)T + log(beta)(1-y)) (Eq. 2).
    The viscosity field is an input; the model is not tested on other rheologies.
  • domain assumption Velocities at time t are fully determined by the instantaneous temperature and viscosity (no memory).
    This follows from the negligible-inertia Stokes limit and justifies the CNN as a static mapping; it would break for non-Newtonian or inertial flows.
  • ad hoc to paper The velocity scaler in Eq. (4) is a valid normalization across the parameter space.
    Fitted to the training data maximum velocities; extrapolation relies on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-based machine learning for mantle convection simulations." pith.science (2026). https://pith.science/paper/Y5KXAXV4

@misc{pith2026250516041,
  author       = {Pith},
  title        = {Pith review of: Physics-based machine learning for mantle convection simulations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y5KXAXV4}},
  note         = {Machine review of arXiv:2505.16041}
}
read the original abstract

Mantle convection simulations are an essential tool for understanding how rocky planets evolve. However, the poorly known input parameters to these simulations, the non-linear dependence of transport properties on pressure and temperature, and the long integration times in excess of several billion years all pose a computational challenge for numerical solvers. We propose a physics-based machine learning approach that predicts creeping flow velocities as a function of temperature while conserving mass, thereby bypassing the numerical solution of the Stokes problem. A finite-volume solver then uses the predicted velocities to advect and diffuse the temperature field to the next time-step, enabling autoregressive rollout at inference. For training, our model requires temperature-velocity snapshots from a handful of simulations (94). We consider mantle convection in a two-dimensional rectangular box with basal and internal heating, pressure- and temperature-dependent viscosity. Overall, our model is up to 89 times faster than the numerical solver. We also show the importance of different components in our convolutional neural network architecture such as mass conservation, learned paddings on the boundaries, and loss scaling for the overall rollout performance. Finally, we test our approach on unseen scenarios to demonstrate some of its strengths and weaknesses.

Figures

Figures reproduced from arXiv: 2505.16041 by the authors.

Figure 1
Figure 1. FIG. 1. The ranges of different fields in the simulations (a-e). Scaling the velocities brings the ranges across [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Our hybrid physics-based model for time-stepping. (a) We use a data-driven approach to model [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Examples of velocity predictions compared to ground truth from the direct solver for two different [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (a) U-net schematic for predicting from state at time [PITH_FULL_IMAGE:figures/full_fig_p018_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Evaluation of rollout performance on three qualitatively different simulations (a,b), (c,d) and (e,f). [PITH_FULL_IMAGE:figures/full_fig_p021_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Parameter space of the training and test sets. We achieve stable rollout on all the simulations in the [PITH_FULL_IMAGE:figures/full_fig_p022_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Comparison of accuracy and speedup of (a) direct solver with 100 momentum skips on the CPU, (b) [PITH_FULL_IMAGE:figures/full_fig_p023_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Results of different ablations, where we change one component of the model at a time. The cross [PITH_FULL_IMAGE:figures/full_fig_p024_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Performance of the baseline and the ML model on three unseen scenarios: (a,b) Extended Boussi [PITH_FULL_IMAGE:figures/full_fig_p028_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

63 extracted references · 45 canonical work pages

  1. [1]

    merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aapmrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  2. [2]

    merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipauth4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translat...

    2010

  3. [3]

    merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs aipnum4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  4. [4]

    merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs apsrev4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  5. [5]

    merlin.mbs apsrmp4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked

    FUNCTION id.bst "merlin.mbs apsrmp4-1.bst 2010-07-25 4.21a (PWD, AO, DPC) hacked" ENTRY address archive archivePrefix author bookaddress booktitle chapter collaboration doi edition editor eid eprint howpublished institution isbn issn journal key language month note number organization pages primaryClass publisher school SLACcitation series title translati...

    2010

  6. [6]

    , author Tosi , N

    author author Agarwal , S. , author Tosi , N. , author Breuer , D. , author Padovan , S. , author Kessel , P. , \ and\ author Montavon , G. ,\ title title A machine-learning-based surrogate model of Mars’ thermal evolution , \ 10.1093/gji/ggaa234 journal journal Geophysical Journal International \ volume 222 ,\ pages 1656--1670 ( year 2020 ) NoStop

    work page doi:10.1093/gji/ggaa234

  7. [7]

    , author Tosi , N

    author author Agarwal , S. , author Tosi , N. , author Hüttig , C. , author Greenberg , D. , \ and\ author Bekar , A. ,\ 10.5281/zenodo.15088589 title PBML\_Mantle\_Convection [Dataset] , \ ( year 2025 ) NoStop

    work page doi:10.5281/zenodo.15088589 2025

  8. [8]

    Accelerating the discovery of steady-states of planetary interior dynamics with machine learning

    author author Agarwal , S. , author Tosi , N. , author Hüttig , C. , author Greenberg , D. S. , \ and\ author Bekar , A. C. ,\ https://arxiv.org/abs/2408.17298 title Accelerating the discovery of steady-states of planetary interior dynamics with machine learning , \ ( year 2024 ),\ http://arxiv.org/abs/2408.17298 arXiv:2408.17298 [physics.flu-dyn] NoStop

    work page Pith review arXiv 2024

Show all 63 references
  1. [9]

    , author Tosi , N

    author author Agarwal , S. , author Tosi , N. , author Kessel , P. , author Breuer , D. , \ and\ author Montavon , G. ,\ title title Deep learning for surrogate modeling of two-dimensional mantle convection , \ 10.1103/PhysRevFluids.6.113801 journal journal Phys. Rev. Fluids \...

    doi

  2. [10]

    , author Pinto , W

    author author Alguacil , A. , author Pinto , W. G. , author Bauerheim , M. , author Jacob , M. C. , \ and\ author Moreau , S. ,\ title title Effects of boundary conditions in fully convolutional networks for learning spatio-temporal dynamics , \ in\ @noop booktitle Machine Lea...

    2021

  3. [11]

    , author Hoyer , S

    author author Alieva , A. , author Hoyer , S. , author Brenner , M. , author Iaccarino , G. , \ and\ author Norgaard , P. ,\ title title Toward accelerated data-driven rayleigh--b \'e nard convection simulations , \ 10.1140/epje/s10189-023-00302-w journal journal The European ...

    doi

  4. [12]

    , author Buttari , A

    author author Amestoy , P. , author Buttari , A. , author L'Excellent , J.-Y. , \ and\ author Mary , T. ,\ title title Performance and Scalability of the Block Low-Rank Multifrontal Factorization on Multicore Architectures , \ @noop journal journal ACM Transactions on Mathemat...

    2019

  5. [13]

    , author Duff , I

    author author Amestoy , P. , author Duff , I. S. , author Koster , J. , \ and\ author L'Excellent , J.-Y. ,\ title title A fully asynchronous multifrontal solver using distributed dynamic scheduling , \ @noop journal journal SIAM Journal on Matrix Analysis and Applications \ v...

    2001

  6. [14]

    , author Dannberg , J

    author author Bangerth , W. , author Dannberg , J. , author Fraters , M. , author Gassmoeller , R. , author Glerum , A. , author Heister , T. , author Myhill , R. , \ and\ author Naliboff , J. ,\ 10.5281/zenodo.14371679 title Aspect v3.0.0 , \ ( year 2024 ) NoStop

    doi

  7. [15]

    \ and\ author Moore , W

    author author Breuer , D. \ and\ author Moore , W. ,\ title title Dynamics and thermal history of the terrestrial planets, the moon, and io , \ in\ 10.1016/B978-0-444-53802-4.00173-1 booktitle Treatise on Geophysics (Second Edition) ,\ Vol. volume 10 ,\ editor edited by\ edito...

    doi

  8. [16]

    , author Marwah , T

    author author Buitrago , R. , author Marwah , T. , author Gu , A. , \ and\ author Risteski , A. ,\ title title On the benefits of memory for modeling time-dependent PDE s , \ in\ https://openreview.net/forum?id=o9kqa5K3tB booktitle The Thirteenth International Conference on Le...

    2025

  9. [17]

    , author Agarwal , S

    author author Catalani , G. , author Agarwal , S. , author Bertrand , X. , author Tost , F. , author Bauerheim , M. , \ and\ author Morlier , J. ,\ title title Neural fields for rapid aircraft aerodynamics simulations , \ 10.1038/s41598-024-76983-w journal journal Scientific R...

    doi

  10. [18]

    , author Di Cola , V

    author author Cuomo , S. , author Di Cola , V. S. , author Giampaolo , F. , author Rozza , G. , author Raissi , M. , \ and\ author Piccialli , F. ,\ title title Scientific Machine Learning Through Physics – Informed Neural Networks : Where we are and What ’s Next , \ 10.1007/s...

    doi

  11. [19]

    \ and\ author Sotin , C

    author author Deschamps , F. \ and\ author Sotin , C. ,\ title title Thermal convection in the outer shell of large icy satellites , \ @noop journal journal Journal of Geophysical Research - Planets \ volume 106 ,\ pages 5107--5121 ( year 2001 ) NoStop

    2001

  12. [20]

    , author Samuel , H

    author author Drilleau , M. , author Samuel , H. , author Rivoldini , A. , author Panning , M. , \ and\ author Lognonné , P. ,\ title title Bayesian inversion of the Martian structure using geodynamic constraints , \ 10.1093/gji/ggab105 journal journal Geophysical Journal Inte...

    doi

  13. [21]

    , author Doin , M.-P

    author author Dumoulin , C. , author Doin , M.-P. , \ and\ author Fleitout , L. ,\ title title Heat transport in stagnant lid convection with temperature- and pressure-dependent newtonian or non-newtonian rheology , \ 10.1029/1999JB900110 journal journal Journal of Geophysical...

    doi

  14. [22]

    a t zu K \

    author author Grimm , V. , author Heinlein , A. , \ and\ author Klawonn , A. ,\ https://kups.ub.uni-koeln.de/64227/ title A short note on solving partial differential equations using convolutional neural networks , \ type Technical Report \ ( institution Universit \"a t zu K \...

    2022

  15. [23]

    , author Breuer , D

    author author Grott , M. , author Breuer , D. , \ and\ author Laneuville , M. ,\ title title Thermo-chemical evolution and global contraction of mercury , \ 10.1016/j.epsl.2011.04.040 journal journal Earth and Planetary Science Letters \ volume 307 ,\ pages 135 -- 146 ( year 2...

    2011 doi

  16. [24]

    , author Koren , T

    author author Gupta , V. , author Koren , T. , \ and\ author Singer , Y. ,\ title title Shampoo: Preconditioned stochastic tensor optimization , \ in\ @noop booktitle International Conference on Machine Learning \ ( year 2018 ) NoStop

    2018

  17. [25]

    ,\ title title A reassessment of the heat transport by variable viscosity convection with plates and lids , \ 10.1029/GL016i002p00179 journal journal Geophy

    author author Gurnis , M. ,\ title title A reassessment of the heat transport by variable viscosity convection with plates and lids , \ 10.1029/GL016i002p00179 journal journal Geophy. Res. Lett. \ volume 16 ,\ pages 179--182 ( year 1989 ) NoStop

    doi

  18. [26]

    , author Tosi , N

    author author H\"uttig , C. , author Tosi , N. , \ and\ author Moore , W. ,\ title title An improved formulation of the incompressible navier-stokes equations with variable viscosity , \ 10.1016/j.pepi.2013.04.002 journal journal Physics of the Earth and Planetary Interiors \ ...

    2013 doi

  19. [27]

    , author Ritschel , T

    author author Innamorati , C. , author Ritschel , T. , author Weyrich , T. , \ and\ author Mitra , N. J. ,\ title title Learning on the Edge : Investigating Boundary Filters in CNNs , \ 10.1007/s11263-019-01223-y journal journal International Journal of Computer Vision \ volum...

    doi

  20. [28]

    author author King , S. D. , author Lee , C. , author van Keken , P. E. , author Leng , W. , author Zhong , S. , author Tan , E. , author Tosi , N. , \ and\ author Kameyama , M. C. ,\ title title A community benchmark for 2-D Cartesian compressible convection in the Earth's ma...

  21. [29]

    author author Knigge , D. M. , author Wessels , D. , author Valperga , R. , author Papa , S. , author Sonke , J.-J. , author Bekkers , E. J. , \ and\ author Gavves , S. ,\ title title Space-time continuous PDE forecasting using equivariant neural fields , \ in\ https://openrev...

    2024

  22. [30]

    , author Smith , J

    author author Kochkov , D. , author Smith , J. A. , author Alieva , A. , author Wang , Q. , author Brenner , M. P. , \ and\ author Hoyer , S. ,\ title title Machine learning–accelerated computational fluid dynamics , \ 10.1073/pnas.2101784118 journal journal Proceedings of the...

    doi

  23. [31]

    author author Korenaga , J. ,\ title title Scaling of plate tectonic convection with pseudoplastic rheology , \ 10.1029/2010JB007670 journal journal Journal of Geophysical Research: Solid Earth \ volume 115 ( year 2010 ),\ 10.1029/2010JB007670 NoStop

    2010 doi

  24. [32]

    \ and\ author Jordan , T

    author author Korenaga , J. \ and\ author Jordan , T. H. ,\ title title Physics of multiscale convection in earth's mantle: Onset of sublithospheric convection , \ 10.1029/2002JB001760 journal journal Journal of Geophysical Research: Solid Earth \ volume 108 ( year 2003 ),\ 10...

    doi

  25. [33]

    , author Sanchez-Gonzalez , A

    author author Lam , R. , author Sanchez-Gonzalez , A. , author Willson , M. , author Wirnsberger , P. , author Fortunato , M. , author Alet , F. , author Ravuri , S. , author Ewalds , T. , author Eaton-Rosen , Z. , author Hu , W. , author Merose , A. , author Hoyer , S. , auth...

    doi

  26. [34]

    , author Veeling , B

    author author Lippe , P. , author Veeling , B. S. , author Perdikaris , P. , author Turner , R. E. , \ and\ author Brandstetter , J. ,\ title title PDE -refiner: Achieving accurate long rollouts with neural PDE solvers , \ in\ https://openreview.net/forum?id=Qv6468llWS booktit...

    2023

  27. [35]

    , author Chu , M

    author author Liu , Q. , author Chu , M. , \ and\ author Thuerey , N. ,\ title title Con FIG : Towards conflict-free training of physics informed neural networks , \ in\ https://openreview.net/forum?id=APojAzJQiq booktitle The Thirteenth International Conference on Learning Re...

    2025

  28. [36]

    , author Mustafa , M

    author author Pathak , J. , author Mustafa , M. , author Kashinath , K. , author Motheau , E. , author Kurth , T. , \ and\ author Day , M. ,\ https://arxiv.org/abs/2010.00072 title Using machine learning to augment coarse-grid computational fluid dynamics simulations , \ ( yea...

    2010 arXiv

  29. [37]

    , author Perdikaris , P

    author author Raissi , M. , author Perdikaris , P. , \ and\ author Karniadakis , G. E. ,\ title title Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , \ @noop journal journ...

    2019

  30. [38]

    , author Hill , D

    author author Ranade , R. , author Hill , D. C. , author Ghule , L. , \ and\ author Pathak , J. ,\ title title A composable machine-learning approach for steady-state simulations on high-resolution grids , \ in\ https://openreview.net/forum?id=jvFTMD5QTq booktitle Advances in ...

    2022

  31. [39]

    , author Solomatov , V

    author author Reese , C. , author Solomatov , V. , \ and\ author Moresi , L.-N. ,\ title title Heat transport efficiency for stagnant lid convection with dislocation viscosity: Application to Mars and Venus , \ @noop journal journal Journal of Geophysical Research - Planets \ ...

    1998

  32. [40]

    , author Fischer , P

    author author Ronneberger , O. , author Fischer , P. , \ and\ author Brox , T. ,\ title title U-net: Convolutional networks for biomedical image segmentation , \ in\ @noop booktitle International Conference on Medical image computing and computer-assisted intervention \ ( orga...

    2015

  33. [41]

    , author Turcotte , D

    author author Schubert , G. , author Turcotte , D. L. , \ and\ author Olson , P. ,\ 10.1017/CBO9780511612879 title Mantle Convection in the Earth and Planets \ ( publisher Cambridge University Press ,\ year 2001 ) NoStop

    2001 doi

  34. [42]

    , author Tosi , N

    author author Schulz , F. , author Tosi , N. , author Plesa , A.-C. , \ and\ author Breuer , D. ,\ title title Stagnant-lid convection with diffusion and dislocation creep rheology: Influence of a non-evolving grain size , \ 10.1093/gji/ggz417 journal journal Geophysical Journ...

    doi

  35. [43]

    , author Boudec , L

    author author Serrano , L. , author Boudec , L. L. , author Koupa \" , A. K. , author Wang , T. X. , author Yin , Y. , author Vittaut , J.-N. , \ and\ author patrick gallinari , ,\ title title Operator learning with neural fields: Tackling PDE s on general geometries , \ in\ h...

    2023

  36. [44]

    author author Shahnas , M. H. \ and\ author Pysklywec , R. N. ,\ title title Toward a unified model for the thermal state of the planetary mantle: Estimations from mean field deep learning , \ 10.1029/2019EA000881 journal journal Earth and Space Science \ volume 7 ( year 2020 ...

    doi

  37. [45]

    , author Hosseinmardi , S

    author author Shishehbor , M. , author Hosseinmardi , S. , \ and\ author Bostanabad , R. ,\ title title Parametric encoding with attention and convolution mitigate spectral bias of neural partial differential equation solvers , \ 10.1007/s00158-024-03834-7 journal journal Stru...

    doi

  38. [46]

    author author Solomatov , V. S. ,\ title title Scaling of temperature‐ and stress‐dependent viscosity convection , \ 10.1063/1.868624 journal journal Physics of Fluids \ volume 7 ,\ pages 266--274 ( year 1995 ) NoStop

    1995 doi

  39. [47]

    author author Stevenson , D. J. , author Spohn , T. , \ and\ author Schubert , G. ,\ title title Magnetism and thermal evolution of the terrestrial planets , \ 10.1016/0019-1035(83)90241-5 journal journal Icarus \ volume 54 ,\ pages 466--489 ( year 1983 ) NoStop

    doi

  40. [48]

    author author Tackley , P. J. ,\ title title Modelling compressible mantle convection with large viscosity contrasts in a three-dimensional spherical shell using the yin-yang grid , \ 10.1016/j.pepi.2008.08.005 journal journal Physics of the Earth and Planetary Interiors \ vol...

    2008 doi

  41. [49]

    , author Breuer , D

    author author Thiriet , M. , author Breuer , D. , author Michaut , C. , \ and\ author Plesa , A.-C. ,\ title title Scaling laws of convection for cooling planets in a stagnant lid regime , \ 10.1016/j.pepi.2018.11.003 journal journal Physics of the Earth and Planetary Interior...

    2018 doi

  42. [50]

    , author Schlachter , K

    author author Tompson , J. , author Schlachter , K. , author Sprechmann , P. , \ and\ author Perlin , K. ,\ title title Accelerating eulerian fluid simulation with convolutional networks , \ in\ @noop booktitle Proceedings of the 34th International Conference on Machine Learni...

    2017

  43. [51]

    , author Breuer , D

    author author Tosi , N. , author Breuer , D. , \ and\ author Spohn , T. ,\ title Evolution of planetary interiors , \ in\ 10.1016/B978-0-12-415845-0.00009-8 booktitle Treatise on Geophysics (Second Edition) ,\ editor edited by\ editor G. Schubert \ ( publisher Elsevier ,\ year...

    doi

  44. [52]

    \ and\ author Padovan , S

    author author Tosi , N. \ and\ author Padovan , S. ,\ title title Mercury, Moon, Mars: Surface expressions of mantle convection and interior evolution on stagnant-lid bodies , \ in\ 10.1002/9781119528609.ch17 booktitle Mantle Convection and Surface Expressions ,\ editor edited...

    doi

  45. [53]

    , author Brand , R

    author author Um , K. , author Brand , R. , author Yun , , author Fei , , author Holl , P. , \ and\ author Thuerey , N. ,\ https://arxiv.org/abs/2007.00016 title Solver-in-the-loop: Learning from differentiable physics to interact with iterative pde-solvers , \ ( year 2021 ),\...

    2007 arXiv

  46. [54]

    , author Morwani , D

    author author Vyas , N. , author Morwani , D. , author Zhao , R. , author Shapira , I. , author Brandfonbrener , D. , author Janson , L. , \ and\ author Kakade , S. M. ,\ title title SOAP : Improving and stabilizing shampoo using adam for language modeling , \ in\ @noop bookti...

    2025

  47. [55]

    , author Schulz , S

    author author Wandel , N. , author Schulz , S. , \ and\ author Klein , R. ,\ title title Metamizer: A versatile neural optimizer for fast and accurate physics simulations , \ in\ https://openreview.net/forum?id=60TXv9Xif5 booktitle The Thirteenth International Conference on Le...

    2025

  48. [56]

    , author Weinmann , M

    author author Wandel , N. , author Weinmann , M. , \ and\ author Klein , R. ,\ title title Learning incompressible fluid dynamics from scratch-towards fast, differentiable fluid models that generalize , \ in\ @noop booktitle International Conference on Learning Representations...

    2020

  49. [57]

    , author Ren , P

    author author Wang , Q. , author Ren , P. , author Zhou , H. , author Liu , X.-Y. , author Deng , Z. , author Zhang , Y. , author Chengze , R. , author Liu , H. , author Wang , Z. , author Wang , J.-X. , author Wen , J.-R. , author Sun , H. , \ and\ author Liu , Y. ,\ title ti...

    2024

  50. [58]

    , author Yao , Y

    author author Wang , Y. , author Yao , Y. , author Guo , J. , \ and\ author Gao , Z. ,\ title title A practical pinn framework for multi-scale problems with multi-magnitude loss terms , \ https://doi.org/10.1016/j.jcp.2024.113112 journal journal Journal of Computational Physic...

    2024

  51. [59]

    , author Langer , S

    author author Wassing , S. , author Langer , S. , \ and\ author Bekemeyer , P. ,\ title title Adopting computational fluid dynamics concepts for physics-informed neural networks , \ in\ 10.2514/6.2025-0269 booktitle AIAA SciTech 2025 Forum \ ( publisher American Institute of A...

    2025 doi

  52. [60]

    \ and\ author Freris , N

    author author Wei , L. \ and\ author Freris , N. M. ,\ title title Multi-scale graph neural network for physics-informed fluid simulation: Multi-scale graph neural network for physics-informed fluid simulation , \ 10.1007/s00371-024-03402-6 journal journal Vis. Comput. \ volum...

    doi

  53. [61]

    , author Maruyama , T

    author author Wu , T. , author Maruyama , T. , \ and\ author Leskovec , J. ,\ https://arxiv.org/abs/2206.07681 title Learning to accelerate partial differential equations via latent global evolution , \ ( year 2022 ),\ http://arxiv.org/abs/2206.07681 arXiv:2206.07681 [cs.LG] NoStop

    2022 arXiv

  54. [62]

    , author Kirchmeyer , M

    author author Yin , Y. , author Kirchmeyer , M. , author Franceschi , J.-Y. , author Rakotomamonjy , A. , \ and\ author Gallinari , P. ,\ https://arxiv.org/abs/2209.14855 title Continuous pde dynamics forecasting with implicit neural representations , \ ( year 2023 ),\ http://...

    2023 arXiv

  55. [63]

    , author McNamara , A

    author author Zhong , S. , author McNamara , A. , author Tan , E. , author Moresi , L. , \ and\ author Gurnis , M. ,\ title title A benchmark study on mantle convection in a 3-d spherical shell using citcoms , \ @noop journal journal Geochemistry, Geophysics, Geosystems \ volu...

    2008

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.