Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Physical Informed Neural Networks for modeling ocean pollutant

T0 review · 4 major / 6 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 This paper sets out to show that a physics-informed neural network with a hybrid, weighted loss can solve the 2D advection-diffusion equation for ocean pollutant transport, and that a tuned 9-layer, 128-neuron configuration reaches about…

desk verdict A modest PINN hyperparameter sweep for 2D advection-diffusion whose 8.25% headline error is an in-sample fit, not an independent accuracy estimate. read the letter →

arxiv 2507.08834 v1 pith:LZPFW62B submitted 2025-07-07 cs.LG

Karishma Battina , Prathamesh Dinesh Joshi , Raj Abhijit Dandekar , Rajat Dandekar , Sreedath Panat This is my paper
classification cs.LG
keywords physics-informedneuralnetworksadvection-diffusionequationoceanpollutantdispersionhybridlossfunctionhyperparametertuningfinitedifferencebenchmarknoisysyntheticdatasurrogatemodeling
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 establish that physics-informed neural networks—networks trained to satisfy the governing PDE as part of their loss—are a workable alternative to grid-based solvers for ocean pollutant transport. It formulates a 2D advection-diffusion equation with a sharp Gaussian initial spill, adds Gaussian noise to a finite-difference solution to mimic field measurements, and trains a network with a hybrid loss that balances physics residuals, boundary/initial conditions, and noisy data. The quantitative centerpiece is that a 9-hidden-layer, 128-neuron network trained with ADAM at learning rate 0.006 for 6,000 iterations reaches approximately 8.25% relative L2 error against the FDM reference, with full-field inference taking about 0.024 seconds. If this holds, PINNs could serve as fast surrogates for pollution forecasting in settings where traditional solvers must be re-run for every parameter change.

What carries the argument

The load-bearing mechanism is the hybrid loss function $L_{\text{total}}(\theta)=L_{\text{physics}}(\theta)+w_{\text{IC}}L_{\text{IC}}(\theta)+w_{\text{data}}L_{\text{data}}(\theta)$, where the physics term is the mean squared PDE residual (with boundary and initial residuals) evaluated at collocation points, $L_{\text{IC}}$ is an explicitly weighted fit to the clean initial Gaussian peak with $w_{\text{IC}}=500$, and $L_{\text{data}}$ is a weighted fit to the noisy FDM data with $w_{\text{data}}=10$. The network is a fully connected feedforward net with tanh activations and three inputs $(t,x,y)$, and automatic differentiation supplies the derivatives in the residual for $\partial_t u + v_x\partial_x u + v_y\partial_y u = D(\partial_{xx}u+\partial_{yy}u)$, with $v_x=v_y=0.5$ and $D=0.01$. This specific loss balance, rather than the network alone, carries the reported accuracy; the strongly weighted initial-condition term anchors the sharp release, and the weighted data term pulls the solution toward the noisy reference while the physics term keeps it consistent with the PDE.

What would settle it

Compute an independent reference for the constant-coefficient problem—an analytic Gaussian solution or a finer-grid, higher-order finite-difference run—and compare the best PINN configuration on time steps and grid points never shown during training; a relative L2 error that rises well above 8.25% on that held-out comparison would show the reported number is an in-sample fit.

Watch

Extended reading notes

Core claim

The central claim is that a PINN can solve the 2D advection-diffusion equation for pollutant dispersion when the loss combines PDE residuals, boundary and initial condition conformity, and weighted fitting to noisy synthetic data, and that this hybrid approach is accurate enough to benchmark against a finite difference solution. The finding is not a new equation but a configuration: architecture depth and width, optimizer, learning rate, and iteration count interact strongly with such a loss. In the sweep, the best run used 9 hidden layers, 128 neurons per layer, ADAM, a learning rate of 0.006, and 6,000 iterations, producing a relative L2 error of about 0.0825 at the final time; the same architecture at 10,000 iterations produced the lowest final loss. The paper also reports that the wider 256-neuron network was less accurate and far slower, and that L-BFGS refinement did not consistently beat a well-tuned ADAM run.

Load-bearing premise

The accuracy claim rests on treating the finite-difference solution on the same 51×51×101 grid that generated the noisy training data as ground truth, with no described held-out validation or convergence check.

Editorial extensions

If this is right

  • At the best configuration, the PINN reaches a relative L2 error of about 8.25% against the FDM reference, the lowest error in the hyperparameter sweep.
  • The 128-neuron architecture beats both the 64-neuron and 256-neuron variants, indicating a capacity sweet spot where adding width increases cost without improving accuracy.
  • The heavily weighted initial-condition loss ($w_{\text{IC}}=500$) consistently anchors the sharp Gaussian release at $t=0$ across scenarios.
  • Once trained, the PINN predicts a full 51×51 field in about 0.024–0.029 seconds, compared with roughly 0.26 seconds for FDM data generation and about 1,168 seconds for the best training run.
  • The hybrid loss with physics, boundary/initial condition, and data terms produces plausible solutions even with 0.5% noise on the data and initial condition and 1% noise on the boundary conditions.

Reading between the lines

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

  • The reported 8.25% error is measured against the same finite-difference solution used to create the noisy training data, with no described held-out split; a separate reference run would be needed before treating that number as independent accuracy.
  • If the accuracy holds, the roughly 0.024-second inference cost makes the trained network a practical surrogate for repeated spill-scenario queries, since the finite difference solver would have to re-solve from scratch for each change.
  • Because the paper uses a constant velocity field, the results do not yet cover spatially or temporally varying currents; testing the same hybrid loss on heterogeneous flows is the natural next step toward real ocean conditions.
  • A clean-data control run, without the added Gaussian noise, would help separate approximation error from noise-handling behavior in the reported L2 numbers.
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

4 major / 6 minor

Summary. The manuscript develops a Physics-Informed Neural Network (PINN) framework in Julia for the 2D advection-diffusion equation as a model of ocean pollutant transport. Training data are generated on a 51×51×101 grid with a finite difference method (FDM), then corrupted with Gaussian noise; the PINN is trained with a hybrid loss combining PDE residuals, boundary/initial condition terms, and a weighted data-fidelity term. The authors perform a hyperparameter sweep over architecture width, optimizer (ADAM, ADAMW, ADAM+LBFGS), learning rate, and iteration count, and report that the best configuration (9 layers, 128 neurons, ADAM, learning rate 0.006, 6000 iterations) achieves a relative L2 error of about 8.25% against the clean FDM solution. The central claim is that this PINN 'successfully developed and benchmarked' solves the 2D advection-diffusion equation for ocean pollution.

Significance. If the 8.25% error were an independent accuracy estimate, the paper would be a useful, if modest, demonstration of PINN hyperparameter tuning for a simple linear PDE with noisy data. The systematic sweep of architectures and optimizers, the explicit hybrid loss formulation, and the use of the Julia scientific computing stack are commendable and reproducible in principle. However, the significance of the main quantitative claim is currently limited because the evaluation is performed on the same grid that supplied the training data, with no held-out split, no repeated runs, and no independent reference solution. As presented, the paper does not yet establish that the proposed PINN generalizes beyond fitting its training data, which is the load-bearing point of the claimed benchmark.

major comments (4)
  1. [§2.2.3, §2.2.4, §4.1, Eq. (9)] The reported 8.25% relative L2 error is an in-sample fit, not an independent accuracy estimate. Training data are generated on the 51×51×101 FDM grid (§2.2.3), and the data-fidelity loss is computed on mini-batches randomly sampled from that same spatio-temporal grid, including t=Tfinal (§2.2.4). Equation (9) in §4.1 then compares the PINN prediction at Tfinal against the clean FDM solution on the identical grid. Because the evaluation points are a subset of the training data, the metric can only measure how well the network memorized the noisy training data, not its predictive accuracy. The paper never describes a held-out split or a separate validation run. The authors should either report an error on a held-out set (e.g., a different random sub-grid, a different noise realization, or a later time) or clearly frame the result as training error.
  2. [§2.2.3, §4.1, Table 2] The FDM solution used as 'ground truth' is not validated. For the constant-coefficient advection-diffusion equation with a Gaussian initial condition on a square domain, an analytic solution is available; comparing the FDM reference against it, or at least performing a grid-convergence study (e.g., 51² vs 101² vs 201²), would establish that the reference is not dominated by discretization error. Without this, the reported errors could reflect error in the reference solution as much as error in the PINN.
  3. [§3.2, Table 2] Each scenario is apparently run only once, with no seed information, no repeated trials, and no error bars. The paper's conclusion that a particular configuration is 'optimal' is based on a single run per hyperparameter setting, so the claimed ranking of configurations (e.g., 9L-128N over 9L-64N and 9L-256N) may not be robust. The authors should report at least a few repeated runs with different random seeds, or clearly state that the results are single-run observations rather than statistically supported comparisons.
  4. [§3.3.3, §3.3.6, Table 2] The narrative contains internal inconsistencies that undermine the benchmark claims. In §3.3.3, Scenario 6 (LR 0.005) is described as achieving 'the lowest final loss in this sweep, 2.5142 × 10^0', but Scenario 9 (LR 0.006, 10,000 iterations) reports a lower final loss of 1.2614 × 10^0 in Table 2. In §3.3.6, the text says 'the fastest training runs (e.g., ADAMW with 1,000 iterations in Scenario 8)' were completed in about 50 s, but Table 2 lists Scenario 8 as ADAM with 6,000 iterations, and no row for ADAMW with 1,000 iterations exists. These discrepancies need to be corrected or reconciled before the reported ranking can be taken at face value.
minor comments (6)
  1. [Title] The title says 'Physical Informed Neural Networks'; it should be 'Physics-Informed Neural Networks'.
  2. [Abstract] The abstract mentions 'non-linear dynamics', but the governing equation (Eq. 1) is linear in u. Rephrase to avoid overstatement.
  3. [§1.1, §1.2] The key contributions are stated nearly identically at the end of §1.1 and §1.2; the duplication should be removed.
  4. [§3.3.3] There is a placeholder reference 'Figure Z' in the description of Scenario 8; replace with the actual figure number.
  5. [Figures 3 and 4] The captions are placeholders ('Caption 1', 'Caption 2'); they should describe the plotted quantities.
  6. [General] The manuscript does not mention code or data availability, which would help reproducibility of the Julia/NeuralPDE-based experiments.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline 8.25% relative L2 error is an in-sample fit: the data-fidelity loss trains on noisy FDM data sampled over the whole spatio-temporal grid (including Tfinal), then Eq. 9 evaluates the same grid points against the clean FDM solution, with no held-out split described.

  1. fitted input called prediction [Sec. 2.2.4 (Data Fidelity Loss and mini-batch sampling) with Sec. 4.1 Eq. (9); data grid defined in Sec. 2.2.3]
    "Data Fidelity Loss (Ldata): This term encourages the network to fit the noisy FDM data over the entire spatio-temporal domain. It is defined as the MSE between the PINN’s predictions and the noisy FDM data, weighted by wdata = 10.0. ... Both Lic and Ldata are computed using mini-batches of size 1,024, randomly sampled at each training iteration. ... The primary metric for accuracy is the Relative L2 Error, which quantifies the normalized difference between the PINN’s predicted solution upred and the clean FDM reference solution uref at the final time step Tfinal = 0.25."

    The accuracy metric is not an independent prediction. The data-fidelity term directly trains the network to minimize MSE against noisy FDM values over the entire spatio-temporal domain, and mini-batches are randomly sampled at each iteration from that same 51x51x101 grid, which includes the final time Tfinal. The reported relative L2 error (Eq. 9) then compares the PINN prediction at Tfinal with the clean FDM solution on the identical grid points that entered training. Because no held-out split or separate validation run is described, the 8.25% figure measures in-sample fit to the training reference rather than generalization or independent solution accuracy.

full rationale

The paper's central claim is that a PINN 'successfully developed and benchmarked ... capable of solving the 2D advection-diffusion equation for modeling ocean pollution,' with an optimal relative L2 error of approximately 8.25%. That claim rests on comparing the trained PINN at Tfinal against the clean FDM solution generated on the same 51x51x101 grid that supplied the noisy training data. Since Ldata is the MSE between the PINN and noisy FDM data over the entire spatio-temporal domain, and mini-batches are randomly sampled from that entire grid (including the final time), the evaluation points are a subset of the training data. The paper does not describe a held-out split, a separate clean simulation for validation, or a convergence study of the FDM reference, and no analytic solution is used even though one is available for this constant-coefficient square-domain problem. Therefore the 8.25% error is an in-sample fit, not an independent estimate of predictive accuracy. This is a clear case of a fitted input being called a prediction, though the PDE residual and the systematic hyperparameter study give the paper some independent content.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central numerical claim relies on FDM-as-ground-truth and on hand-set loss weights and noise levels. No new physical entities are introduced. Because the evaluation is in-sample, the headline error should be read as a fit quality measure rather than a predictive claim.

free parameters (6)
  • Initial condition loss weight wIC = 500.0
    Hand-set to force IC matching; central to the hybrid loss in Sec 2.2.4 and not justified by a sweep.
  • Data fidelity loss weight wdata = 10.0
    Hand-set weighting of the noisy FDM fit in Sec 2.2.4; no grid search or sensitivity analysis is shown.
  • Learning rate = 0.006
    Selected as the best value from the sweep, on the same metric used for the final claim (Table 2).
  • Neural network architecture = 9 layers, 128 neurons
    Chosen post hoc because it produced the lowest L2 error in Table 2; no independent validation.
  • Data noise standard deviation sigma_data = 0.005 * std(Uclean)
    Ad hoc synthetic noise level in Sec 2.2.3; no real observation noise calibration.
  • Boundary/initial noise levels = sigma_BC = 0.01, sigma_IC = 0.5%
    Ad hoc choices in Sec 2.2.1 to simulate measurement error; no real data basis.
assumptions (5)
  • domain assumption The FDM solution on the 51x51x101 grid is sufficiently accurate to serve as ground truth.
    All training labels and the benchmark Eq. 9 use this same grid; no convergence study or analytic validation is provided (Sec 2.2.3, Sec 4.1).
  • domain assumption The 2D advection-diffusion equation with constant v and D models the intended ocean pollutant transport.
    The ocean relevance claim assumes this PDE captures the phenomenon despite constant velocity and a unit-square domain (Sec 2.1, Sec 5).
  • ad hoc to paper Gaussian additive noise is representative of real ocean measurement error.
    No real data is used; noise is generated synthetically to make the problem look realistic (Sec 2.2.3).
  • domain assumption NeuralPDE.jl symbolic residual construction correctly encodes the PDE, initial, and boundary conditions.
    The physics loss is not hand-verified; the paper relies on library correctness (Sec 2.2.4).
  • ad hoc to paper Single optimizer runs are sufficient to characterize performance.
    No seeds or repeated runs are reported, yet conclusions are drawn from one run per configuration (Sec 3.3.5).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physical Informed Neural Networks for modeling ocean pollutant." pith.science (2026). https://pith.science/paper/LZPFW62B

@misc{pith2026250708834,
  author       = {Pith},
  title        = {Pith review of: Physical Informed Neural Networks for modeling ocean pollutant},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LZPFW62B}},
  note         = {Machine review of arXiv:2507.08834}
}
read the original abstract

Traditional numerical methods often struggle with the complexity and scale of modeling pollutant transport across vast and dynamic oceanic domains. This paper introduces a Physics-Informed Neural Network (PINN) framework to simulate the dispersion of pollutants governed by the 2D advection-diffusion equation. The model achieves physically consistent predictions by embedding physical laws and fitting to noisy synthetic data, generated via a finite difference method (FDM), directly into the neural network training process. This approach addresses challenges such as non-linear dynamics and the enforcement of boundary and initial conditions. Synthetic data sets, augmented with varying noise levels, are used to capture real-world variability. The training incorporates a hybrid loss function including PDE residuals, boundary/initial condition conformity, and a weighted data fit term. The approach takes advantage of the Julia language scientific computing ecosystem for high-performance simulations, offering a scalable and flexible alternative to traditional solvers

Figures

Figures reproduced from arXiv: 2507.08834 by the authors.

Figure 1
Figure 1. Solution plots and Adam loss history for Scenario 1 (9L-64N network, 1000 iterations) [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Solution plots and Adam and LBFGS loss history for Scenario 2 (9L-64N network, 3000 iterations) [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Solution plots and Adam loss history for Scenario 3 (9L-64N network, 10,000 Adam iterations) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Solution plots and AdamW loss history for Scenario 4 (9L-64N network, 10,000 Adam iterations) [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Solution plots and AdamW loss history for Scenario 5 (9L-128N network, 6000 Adam iterations) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Solution plots and Adam loss history for Scenario 6 (9L-128N network, 6000 Adam iterations, 0.005 LR) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Solution plots and Adam loss history for Scenario 8 (9L-128N network, 6000 Adam iterations, 0.006 LR) [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]
Figure 8
Figure 8. Figure 8: Solution plots and Adam loss history for Scenario 9 (9L-128N network, 10000 Adam iterations, 0.006 LR) [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Solution plots and Adam loss history for Scenario 10 (9L-256N network, 6000 Adam iterations, 0.006 LR) [PITH_FULL_IMAGE:figures/full_fig_p010_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Mass-Conserving Physics-Informed Neural Networks For The One-Dimensional Advection-Diffusion Equation

    physics.comp-ph 2026-07 conditional novelty 3.0 of 10

    Adding a soft mass-conservation penalty to PINNs for the 1D advection-diffusion equation reduces long-term relative L2 error by 9–67× and mass error by 15–215× compared to vanilla PINNs across Peclet numbers 0.01–20.

Reference graph

Works this paper leans on

30 extracted references · 23 canonical work pages · cited by 1 Pith paper

  1. [1]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George Em Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. Journal of Computational Physics, 378:686–707, 2019

    2019

  2. [2]

    DeepXDE: A deep learning library for solving differential equations

    Lu Lu, Xuhui Meng, Zhiping Mao, and George Em Karniadakis. DeepXDE: A deep learning library for solving differential equations. SIAM Review, 63(1):208–228, 2021

    2021

  3. [3]

    When and why PINNs fail to train: A neural tangent kernel perspective

    Shunjin Wang, Xuhui Yu, and Paris Perdikaris. When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 449:110768, 2022

    work page 2022

  4. [4]

    Effective training of PINNs by combining CMA-ES with gradient descent

    Yujie Yang, Paris Perdikaris, and Weinan E. Effective training of PINNs by combining CMA-ES with gradient descent. Computers & Mathematics with Applications , 127:27–41, 2024

    work page 2024

  5. [5]

    Neural tangent kernel analysis of PINN for advection- diffusion equation

    Mohammad Hossein Saadat and Mohammad Amin Ghatei. Neural tangent kernel analysis of PINN for advection- diffusion equation. arXiv preprint arXiv:2211.11716, 2022

    arXiv 2022

  6. [6]

    Physics Informed Neural Networks with strong and weak residuals for advection-dominated diffusion problems

    Maciej Paszy´nski, Bartosz Sawicki, Michał Krzemi´nski, and Szymon Draga. Physics Informed Neural Networks with strong and weak residuals for advection-dominated diffusion problems. arXiv preprint arXiv:2307.07647, 2023

    work page Pith review arXiv 2023

  7. [7]

    Long-term simulation of physical and mechanical behaviors using curriculum-transfer-learning based physics-informed neural networks

    Zhuojia Fu, Yuxin Zhang, and Yifan Wang. Long-term simulation of physical and mechanical behaviors using curriculum-transfer-learning based physics-informed neural networks. arXiv preprint arXiv:2502.07325, 2023

    arXiv 2023

  8. [8]

    S., Giampaolo, F., Rozza, G., Raissi, M., & Piccialli, F

    Cuomo, S., Di Cola, V . S., Giampaolo, F., Rozza, G., Raissi, M., & Piccialli, F. (2022). Scientific machine learning through physics-informed neural networks: A review. Journal of Computing, 104(2), 447–466

    work page 2022

Show all 30 references
  1. [9]

    E., Kevrekidis, I

    Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., & Yang, L. (2021). Physics-informed machine learning. Nature Reviews Physics, 3, 422–440

    2021

  2. [10]

    Zhang, Y ., & Baptista, A. M. (2008). An efficient and robust particle-tracking model for coastal and estuarine water systems. Environmental Modelling & Software, 23(1), 41–52

    2008

  3. [11]

    Okubo, A. (1971). Oceanic diffusion diagrams. Deep Sea Research and Oceanographic Abstracts, 18(8), 789–802

    1971

  4. [12]

    Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., Carvalhais, N., & Prabhat. (2019). Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743), 195–204

    2019

  5. [13]

    S., & Perdikaris, P

    Zhu, Y ., Zabaras, N., Koutsourelakis, P. S., & Perdikaris, P. (2019). Physics-constrained deep learning for high- dimensional surrogate modeling and uncertainty quantification without labeled data. Journal of Computational Physics, 394, 56–81

    2019

  6. [14]

    LeVeque, R. J. (2002). Finite V olume Methods for Hyperbolic Problems. Cambridge University Press

    2002

  7. [15]

    D., & O’Brien, J

    Smith, R. D., & O’Brien, J. J. (1983). Numerical simulation of pollutant dispersion in the ocean. Journal of Physical Oceanography, 13(5), 857–869

    1983

  8. [16]

    Bezanson, J., Edelman, A., Karpinski, S., & Shah, V . B. (2017). Julia: A fresh approach to numerical computing. SIAM Review, 59(1), 65–98

    2017

  9. [17]

    Rackauckas, C., Ma, Y ., Martensen, J., Warner, C., Gwozdz, K., & Innes, M. (2020). Universal differential equations for scientific machine learning. arXiv preprint, arXiv:2001.04385. 12 arXiv Template A PREPRINT

    2020 arXiv

  10. [18]

    Sun, L., Gao, H., Pan, S., & Wang, J.-X. (2020). Surrogate modeling for fluid flows based on physics-constrained deep learning without simulation data. Computer Methods in Applied Mechanics and Engineering , 361, 112732

    2020

  11. [19]

    B., List, E

    Fischer, H. B., List, E. J., Koh, R. C. Y ., Imberger, J., & Brooks, N. H. (1979). Mixing in Inland and Coastal Waters. Academic Press

    1979

  12. [20]

    P., & Verron, J

    Chassignet, E. P., & Verron, J. (Eds.). (2006). Ocean Weather Forecasting: An Integrated View of Oceanography. Springer

    2006

  13. [21]

    W., & Mayers, D

    Morton, K. W., & Mayers, D. F. (2005). Numerical Solution of Partial Differential Equations: An Introduction. Cambridge University Press

    2005

  14. [22]

    Shankar, V ., & Wright, G. B. (2022). A survey of mesh-free and physics-informed neural network methods for PDEs. Journal of Computational Physics, 469, 111543

    2022

  15. [23]

    Lguensat, R., Sun, M., Fablet, R., & Chen, G. (2022). Machine learning for oceanography: Challenges and opportunities. Journal of Marine Systems, 230, 103672

    2022

  16. [24]

    Zhu, X., & Zabaras, N. (2018). Bayesian deep convolutional encoder–decoder networks for surrogate modeling and uncertainty quantification. Journal of Computational Physics, 366, 415–447

    2018

  17. [25]

    Rackauckas, C., & Nie, Q. (2017). DifferentialEquations.jl – A performant and feature-rich ecosystem for solving differential equations in Julia. Journal of Open Research Software, 5(1), 15

    2017

  18. [26]

    P.; Ba, J

    Kingma, D. P.; Ba, J. Adam: A Method for Stochastic Optimization. arXiv preprint arXiv:1412.6980, 2014

    2014 arXiv

  19. [27]

    Decoupled Weight Decay Regularization

    Loshchilov, I.; Hutter, F. Decoupled Weight Decay Regularization. arXiv preprint arXiv:1711.05101, 2017

    2017 arXiv

  20. [28]

    Updating quasi-Newton matrices with a limited storage

    Nocedal, J. Updating quasi-Newton matrices with a limited storage. Mathematics of Computation 1980, 35(151), 773–782

    1980

  21. [29]

    A Review of First-Order and Quasi-Newton Methods for Deep Learning

    Schmidt, M. A Review of First-Order and Quasi-Newton Methods for Deep Learning. Technical Report, University of British Columbia, 2020

    2020

  22. [30]

    On the Use of L-BFGS in the Training of Physics-Informed Neural Networks

    Jarlebring, E.; Tichý, P. On the Use of L-BFGS in the Training of Physics-Informed Neural Networks. arXiv preprint arXiv:2201.03921, 2022. 13

    2022 arXiv

Pith tools

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