Pith. sign in

REVIEW 4 major objections 6 minor 49 references

The paper proposes that under a misspecified governing PDE, the true solution operator can be decomposed into a physics-based prior operator and a learnable correction operator, and that training a serial DeepONet with the correction placed

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 ↗

A serial DeepONet framework decomposes a solution operator into a physics-based prior plus a learned correction, reporting reduced errors under misspecified PDEs.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection A promising operator-correction idea undermined by a central mismatch between the claimed solution-space decomposition and the actual losses. the 4 major comments →

arxiv 2606.03469 v2 pith:SYHD4KBT submitted 2026-06-02 math.NA cs.NA

Physics-guided correction for operator learning under model misspecification

Lei Ma , Nicolas Boull\'e , Yu-Sen Yang , Hao Wu , Ling Guo This is my paper
classification math.NA cs.NA MSC 65M9968T07
keywords operator learningmodel misspecificationphysics-informed learningDeepONetcorrection operatorPDE residualuncertainty quantificationserial architecture
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.

The reading

The paper argues that when the governing PDE used in physics-informed operator learning is misspecified, the learned solution operator carries systematic bias, and it proposes to repair this at the operator level: write the true solution map as the sum of a prior operator (the inverse of the approximate PDE) and a learnable correction operator. The correction is trained by embedding it inside the residual of the prior PDE and by minimizing a data-consistency term, using a serial DeepONet in which the second network conditions on both the input function and the prior network's prediction. Across four benchmark problems (diffusion-reaction, Burgers, cavity flow, and hyperelastic beam), the corrected model reduces relative L2 errors by orders of magnitude compared with training under the misspecified PDE alone, and the correction term itself recovers the physical missing term (e.g., -k_r u, u ∂_x u, ν ∂_xx u). The paper also demonstrates tolerance to sparse and noisy observations and shows that deep ensembles provide meaningful uncertainty bands. The main claim is that decomposing prediction into a physics-based prior and a learned discrepancy term lets operator learning retain physical structure while adapting to inaccurate equations.

Core claim

The central claim is that model misspecification in operator learning can be corrected not by changing the PDE but by learning an additive operator on top of a physics-induced prior. Specifically, the paper sets N† ≈ Gθ + Gψ, where Gθ approximates the inverse of a possibly wrong differential operator N0 and Gψ(v, Gθ(v)) approximates the discrepancy N† − N†0. Training minimizes L_physics = ||N0[Gθ(v)] + Gψ(v, uθ) − v||² plus a data loss L_data = ||Gθ(v)(y_obs) − u(y_obs)||². The paper reports that when the prior PDE is misspecified (wrong reaction term, omitted advection or diffusion, Newtonian instead of power-law viscosity, or linear instead of hyperelastic constitutive law), the corrected

What carries the argument

The load-bearing construction is a serial DeepONet: a trunk-branch network Gθ maps the input function v (sampled at sensors) to a prior predicted solution uθ; a second trunk-branch network Gψ takes both v and uθ (sampled on a coarse grid) and produces an additive correction. The correction is inserted into the residual of the prior differential operator N0, so the physics loss becomes N0[uθ] + Gψ(v, uθ) − v, and both networks are trained jointly with a data term on uθ. This design makes the correction state-dependent—it can learn missing nonlinear or derivative terms as functions of the current solution—while the prior operator still imposes the known part of the physics.

Load-bearing premise

The load-bearing premise is that the training loss as written drives the full corrected sum Gθ+Gψ toward the data, even though the data term in Eq. (2) is applied only to Gθ.

What would settle it

Train the exact serial DeepONet with the data loss moved from Gθ alone to Gθ+Gψ (the corrected prediction), keeping all other settings fixed; if the error does not stay at or below the reported levels, then the reported gains depend on supervision that the stated loss function does not provide.

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

If this is right

  • If the decomposition works as intended, physics-informed operator learning no longer requires the prescribed PDE to be exactly right: approximate models become usable priors.
  • The correction operator recovers interpretable missing terms (e.g., -k_r u, u ∂_x u, ν ∂_xx u), so the framework doubles as a data-driven discrepancy-discovery tool.
  • Because the formulation is architecture-independent, the same prior-plus-correction idea extends to other neural operators and even classical numerical solvers as the prior.
  • Sparse and noisy observations are shown to be tolerable: ensemble uncertainty widens but the mean stays close to the reference, suggesting a practical route to uncertainty quantification for imperfect models.
  • In the hyperelasticity benchmark, the corrected model beats a standard data-only DeepONet under the same limited observations, indicating that combining an approximate physics prior with a learned correction extracts more from data than pure regression.

Where Pith is reading between the lines

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

  • The stated training objective leaves the corrected sum Gθ+Gψ unsupervised: L_data only constrains Gθ, so the reported numerical gains rest on the physics residual indirectly pulling the sum toward the data. A direct test is to add the data loss to Gθ+Gψ; if accuracy is maintained or improved, the decomposition is more reliably enforced.
  • The framework can be viewed as a model-discrepancy method at the operator level, conceptually parallel to classical Bayesian calibration—but it learns a functional correction without parameter sampling.
  • A testable extension is to apply the same serial correction to time-dependent problems with parameter drift by letting the correction depend on a time index or latent parameter; the paper's examples are mostly steady or single-parameter, so dynamic drift remains open.
  • The state-dependent conditioning (v, uθ) suggests the correction learns a closure term; one could inspect Gψ to see whether it converges to the analytic missing term as data increase, which would provide a falsifiable check on identifiability.
Share X Bluesky LinkedIn Reddit HN

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 / 6 minor

Summary. The paper proposes a physics-guided operator correction framework for operator learning under model misspecification. The true solution operator is decomposed into a prior operator, induced by an approximate PDE model, and a learnable correction operator; the authors instantiate this with a serial DeepONet in which the first network produces the prior prediction u_θ and the second produces a correction conditioned on both the input function and u_θ. The training objective combines a physics residual built from the prior operator plus the correction, and a data-consistency loss. Numerical experiments on diffusion-reaction, Burgers, cavity flow, and hyperelasticity report that the corrected model substantially reduces errors relative to the misspecified baseline, and additional tests address sparse/noisy observations and deep-ensemble uncertainty.

Significance. The conceptual idea of correcting a prior solution operator rather than re-learning the full map is timely, and the serial DeepONet realization is a natural implementation. The four benchmark problems cover usefully different types of misspecification, and the sparse/noisy observation experiments address practical robustness. However, the central training objective as written does not train the corrected prediction that the paper claims to evaluate: the data loss supervises only G_θ, and the physics loss places G_ψ in residual/source space rather than solution space. Consequently, the reported numerical improvements are not derivable from the stated method. If the formulation is corrected and the experiments re-run, the framework could be a useful contribution; in its current form the central claim is not supported.

major comments (4)
  1. [§2.2, Eq. (2); Algorithm 1, line 5] The data-consistency loss is defined on G_θ(v)(y_obs) alone, and Algorithm 1 confirms this. The final corrected prediction G_θ(v)+G_ψ(v,u_θ) is never compared with the observed solutions u. Thus nothing in the objective enforces the stated target N†[v]≈G_θ(v)+G_ψ(v,u_θ). The numerical results cannot be attributed to the described training procedure. The data loss must be defined on the corrected prediction, or the corrected prediction must be supervised by an additional term.
  2. [§2.2, Eq. (1)] Eq. (1) imposes N_0[G_θ(v)] + G_ψ(v,u_θ) - v = 0. This treats G_ψ as an additive correction to the residual/source space, not as the solution-space discrepancy N†[v]-N_0†[v] defined in §2.1. If G_θ actually approximates the prior solution operator, then N_0[G_θ(v)]≈v and minimizing Eq. (1) drives G_ψ to zero. If instead G_θ is fitted to the true data as in Eq. (2), then G_ψ is forced to absorb the prior-model residual N_0[G_θ]-v, not the solution discrepancy. Either way, the training objective is inconsistent with the claimed decomposition. The authors need to derive Eq. (1) from the decomposition or replace it with a residual that acts on the corrected solution, e.g. N_0[G_θ+G_ψ]-v.
  3. [§3.1 and §3.2] The reported correction targets are inconsistent with the stated misspecification. In the diffusion-reaction example the prior is D∂²u - k_r = v and the true model is D∂²u - k_r u = v, so the residual discrepancy is k_r(1-u), not -k_r u; yet the paper defines ϕ=-k_r u as the correction quantity. In Burgers Case C, the prior omits diffusion, so the correction needed in Eq. (1) is -ν∂²u/∂x², while the text lists +ν∂²u/∂x². These sign/target mismatches mean the experiments do not implement the discrepancy described in §2.1, and the tables cannot be interpreted as validating the proposed correction.
  4. [§2.1, Fig. 1] The schematic in Fig. 1 shows the corrected prediction as G_θ(v) ⊕ G_ψ(v,u_θ), and the caption states that this combination approximates the true operator. But the training diagram in Fig. 2 places the data loss on u_θ only, and the physics loss on N_0[G_θ]+G_ψ. The two figures therefore depict different training objectives. The manuscript needs a single, consistent statement of what is optimized.
minor comments (6)
  1. [Abstract/Introduction] The phrase 'usually requires typically' is redundant; the sentence should be rewritten.
  2. [§3.1] Typo: 'manifactured solutions' should be 'manufactured solutions'.
  3. [§3, Table 1] The Adam hyperparameters are given as β_1=0.999 and β_2=0.999. This is unusual; standard values are β_1=0.9, β_2=0.999. Please clarify or correct.
  4. [§3.1, Table 2] The table header lists 'f' but the text and equations use v for the source term; please unify notation.
  5. [§3.3, Table 4] The pressure error is reported as '0.01±0.01%' and '0.80%±0.04%', which is inconsistent formatting; also clarify whether these are percentages or relative L2 errors.
  6. [General] No code or data availability statement is provided. Given the sensitivity of the experiments to the training objective, releasing code would materially help reproducibility.

Circularity Check

0 steps flagged

No significant circularity: the correction operator is learned from data and the self-citations are not load-bearing; the Eq. (1)-(2) supervision gap is a correctness issue, not a circular reduction.

full rationale

The paper's central decomposition N†[v] = N0†[v] + (N†[v] − N0†[v]) (Section 2.1) is an algebraic identity, but it is used as a modeling ansatz rather than as a derivation of the numerical results; the correction operator Gψ is fitted from data, not derived from the claimed target. The training objective in Eqs. (1)-(2) is inconsistent with the claimed decomposition: Eq. (2) supervises only Gθ(v), and Eq. (1) drives Gψ toward the source-space residual v − N0[Gθ], not toward the solution-space discrepancy N†[v] − N0†[v]. This is a potentially serious correctness gap, but it is not circularity: no equation in the paper makes the reported prediction equal to its input by construction, and no fitted parameter is relabeled as a prediction. The citations to prior work by Boullé et al. ([10, 29-32, 46]) are contextual—related-work pointers and benchmark setup—and are not used to justify the correction framework itself. Numerical benchmarks rely on external datasets and solvers, so the results are not forced by a self-citation chain. Therefore, no circular step is identified.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The method introduces a correction operator as a network component, but no new physical entity, force, particle, or conserved quantity. The key additional structure is the ad hoc residual-correction equation, which is where the paper's main burden lies.

free parameters (3)
  • Loss weights λ_bc, λ_ic, λ_u, λ_e, λ_d = λ_bc=1, λ_ic=λ_u=50 (Burgers); λ_e=100, λ_u=100000, λ_bc=1 (hyperelastic)
    Chosen by hand; no sensitivity study is reported, and the balance between physics and data terms affects accuracy.
  • DeepONet truncation p1=p2 and network width/depth = p1=p2=100; depth 4; widths 64, 128, 256 per problem in Table 1
    Architecture hyperparameters chosen without ablation; the correction operator capacity is matched to the prior network.
  • Observation and collocation counts = e.g., 101 sensors for diffusion-reaction, 51×51 grid for uθ in Burgers, 250 interior velocity points for cavity flow
    Data availability choices affect the reported accuracy; no systematic sensitivity analysis is given.
axioms (3)
  • domain assumption The inverse solution operators N† and N0† exist and are well approximated by neural operators Gθ and Gψ.
    Section 2.1 assumes the decomposition N†[v] = N0†[v] + (N†[v]−N0†[v]) and that both pieces can be represented by DeepONets.
  • ad hoc to paper Adding Gψ to the prior residual N0[Gθ] in Eq. (1) is equivalent to correcting the solution operator.
    No derivation is provided for why N0[Gθ] + Gψ = v should make Gθ+Gψ approximate N†; this is the central unproven step.
  • ad hoc to paper The input function v can be evaluated at spatio-temporal collocation points y_f even when v is defined only in space (e.g., Burgers initial condition).
    Section 3.2 uses v as an initial condition while Eq. (1) evaluates v(y_f) on a 101×101 space-time grid, which is not well-defined as written.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-guided correction for operator learning under model misspecification." pith.science (2026). https://pith.science/paper/SYHD4KBT

@misc{pith2026260603469,
  author       = {Pith},
  title        = {Pith review of: Physics-guided correction for operator learning under model misspecification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SYHD4KBT}},
  note         = {Machine review of arXiv:2606.03469}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Physics-informed operator learning provides an efficient framework for approximating solution operators of partial differential equations by combining observational data with governing physical laws. However, most existing methods implicitly assume that the prescribed governing equation is accurate. This assumption may fail in practical applications, where model simplifications, missing physical effects, parameter drift, or incomplete constitutive relations can lead to model misspecification. In this work, we propose a physics-guided operator correction framework for learning solution operators under misspecified governing equations. At the operator level, the target mapping is decomposed into a prior operator induced by an approximate physical model and a learnable correction operator that accounts for the remaining discrepancy. We realize the operator model correction using a serial DeepONet architecture, where the first DeepONet provides a solution prediction and the second DeepONet learns an additive correction conditioned on both the input function and the solution prediction. The learned correction is incorporated into the physics residual and trained together with data-consistency constraints, allowing the model to retain useful physical structure while adapting to inaccurate governing equations. Numerical experiments on diffusion-reaction, Burgers, cavity flow, and hyperelastic problems show that the proposed method substantially reduces errors induced by misspecified physics. Additional tests under sparse and noisy observations further demonstrate the robustness of the framework and its ability to provide informative uncertainty estimates through deep ensembles.

Figures

Figures reproduced from arXiv: 2606.03469 by Hao Wu, Lei Ma, Ling Guo, Nicolas Boull\'e, Yu-Sen Yang.

Figure 1
Figure 1. Figure 1: Schematic illustration of the unified operator correction framework under model misspecification. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic of the physics-guided serial DeepONet operator correction framework. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: 1d diffusion-reaction equation. (a):Relative L2 errors between corrected model predictions and the reference solution for u (red) and ϕ (blue) as Nu varies, with Nf = 1000 and Np = 1000; (b): Relative L2 error of the source term v versus Np (blue, with Nu = 100 and Nf = 1000) and versus Nf (red, with Nu = 100 and Np = 1000). Shaded regions indicate one standard deviation over five independent experiments. … view at source ↗
Figure 4
Figure 4. Figure 4: 1d diffusion-reaction equation. Reference solution u(x) (black solid line) and prediction (red dashed line) for a representative test sample, (a)-(c) show the results for the known, misspecified, and corrected models, respectively. In [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 1d diffusion-reaction equation under imperfect observations. (a): Prediction of u(x) with five sparse observations; (b): Prediction of u(x) with noisy observations at noise level σ = 0.02. The black solid line denotes the reference solution, the red dashed lines denote the ensemble mean prediction, and the shaded regions indicate two standard deviations estimated from ten independent runs. The square marke… view at source ↗
Figure 6
Figure 6. Figure 6: 1d burgers’ equation. Relative L2 error of the solution u versus the number of solution observations Nu for Cases A–C under the corrected model. The results are obtained with Np = 2000 training samples and an Nf = 101 × 101 collocation grid. Markers and shaded regions denote the mean and one standard deviation, respectively, computed over five independent runs. tional cubic term ϵu 3 [28]: ∂u ∂t + u ∂u ∂x … view at source ↗
Figure 7
Figure 7. Figure 7: 1d burgers’ equation. Correction term ϕ(x, t) for a representative test sample in three cases. Columns 1–3 display the reference solution and the corresponding predictions of the misspecified and corrected models, respectively; column 4 reports the absolute error of the corrected model relative to the reference. deviate substantially from the reference in both magnitude and spatio-temporal distribution. In… view at source ↗
Figure 8
Figure 8. Figure 8: 1d burgers’ equation. Correction term ϕ at t = 0.5 under sparse solution observations for Cases A–C. Results are shown for 100 (top) and 200 (bottom) solution observations. The black line denotes the reference solution, the red dashed line denotes the ensemble mean prediction, and the shaded regions indicates two standard deviation computed from five independent runs. 12 [PITH_FULL_IMAGE:figures/full_fig_… view at source ↗
Figure 9
Figure 9. Figure 9: 2d cavity flow. Representative velocity field results for a test realization, where the top and bottom rows show the x- and y- components ux and uy, respectively. From left to right, the columns report the reference solution, the misspecified prediction, the corrected prediction, and the absolute error of the corrected prediction relative to the reference [PITH_FULL_IMAGE:figures/full_fig_p014_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: 2d hyperelastic problem. Displacement field components for a representative test sample with ϵ = 0.0991: (a) ux and (b) uy. For each component, the first row presents the reference solution and the predictions of the misspecified and corrected models, while the second row shows the corresponding absolute errors relative to the reference solution. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗

Review history (2 revisions) →

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

49 extracted references · 4 linked inside Pith

  1. [1]

    G. K. Batchelor, An introduction to fluid dynamics, Cambridge University Press, 2000

    2000

  2. [2]

    F. P . Incropera, D. P . DeWitt, T. L. Bergman, A. S. Lavine, et al., Fundamentals of heat and mass transfer, Wiley New York, 1996

    1996

  3. [3]

    R. B. Bird, Transport phenomena, Appl. Mech. Rev. 55 (2002) 1–4

    2002

  4. [4]

    J. N. Goodier, P . Hodge Jr, Elasticity and plasticity: The mathematical theory of elasticity and the mathematical theory of plasticity, Dover Publications, 2016

    2016

  5. [5]

    D. E. Keyes, L. C. McInnes, C. Woodward, W. Gropp, E. Myra, M. Pernice, J. Bell, J. Brown, A. Clo, J. Connors, et al., Multiphysics simulations: Challenges and opportunities, Int. J. High Perform. Comput. Appl. 27 (2013) 4–83

    2013

  6. [6]

    Raissi, P

    M. Raissi, P . Perdikaris, G. E. Karniadakis, Physics-informed neural networks: A deep learning frame- work for solving forward and inverse problems involving nonlinear partial differential equations, J. Comput. Phys. 378 (2019) 686–707

    2019

  7. [7]

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

    2021

  8. [8]

    Kovachki, Z

    N. Kovachki, Z. Li, B. Liu, K. Azizzadenesheli, K. Bhattacharya, A. Stuart, A. Anandkumar, Neural operator: Learning maps between function spaces with applications to PDEs, J. Mach. Learn. Res. 24 (2023) 1–97

    2023

  9. [9]

    N. B. Kovachki, S. Lanthaler, A. M. Stuart, Operator learning: Algorithms and analysis, in: Handbook of Numerical Analysis, volume 25, Elsevier, 2024, pp. 419–467

    2024

  10. [10]

    Boullé, A

    N. Boullé, A. Townsend, A mathematical guide to operator learning, in: Handbook of Numerical Analysis, volume 25, Elsevier, 2024, pp. 83–125

    2024

  11. [11]

    L. Lu, P . Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators, Nat. Mach. Int. 3 (2021) 218–229

    2021

  12. [12]

    Z. Li, N. B. Kovachki, K. Azizzadenesheli, B. liu, K. Bhattacharya, A. Stuart, A. Anandkumar, Fourier neural operator for parametric partial differential equations, in: International Conference on Learning Representations, 2021

    2021

  13. [13]

    Haghighat, U

    E. Haghighat, U. bin Waheed, G. Karniadakis, En-deeponet: An enrichment approach for enhancing the expressivity of neural operators with applications to seismology, Comput. Methods Appl. Mech. Eng. 420 (2024) 116681

    2024

  14. [14]

    Y. Qiu, N. Bridges, P . Chen, Derivative-enhanced deep operator network, in: Advances in Neural Information Processing Systems, volume 37, 2024, pp. 20945–20981

    2024

  15. [15]

    A. Tran, A. Mathews, L. Xie, C. S. Ong, Factorized fourier neural operators, in: International Conference on Learning Representations, 2023

    2023

  16. [16]

    Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, A. Stuart, K. Bhattacharya, A. Anandkumar, Multipole graph neural operator for parametric partial differential equations, in: Advances in Neural Information Processing Systems, volume 33, 2020, pp. 6755–6766

    2020

  17. [17]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems, volume 30, 2017. 18

    2017

  18. [18]

    Bommasani, D

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M. S. Bernstein, J. Bohg, A. Bosselut, E. Brunskill, et al., On the opportunities and risks of foundation models, arXiv preprint arXiv:2108.07258 (2021)

    Pith/arXiv arXiv 2021

  19. [19]

    B. Shih, A. Peyvan, Z. Zhang, G. E. Karniadakis, Transformers as neural operators for solutions of differential equations with finite regularity, Comput. Methods Appl. Mech. Eng. 434 (2025) 117560

    2025

  20. [20]

    Z. Hao, Z. Wang, H. Su, C. Ying, Y. Dong, S. Liu, Z. Cheng, J. Song, J. Zhu, Gnot: A general neural operator transformer for operator learning, in: International conference on machine learning, 2023, pp. 12556–12569

    2023

  21. [21]

    J. Sun, Y. Liu, Z. Zhang, H. Schaeffer, Towards a foundation model for partial differential equations: Multioperator learning and extrapolation, Phys. Rev. E 111 (2025) 035304

    2025

  22. [22]

    S. Wang, H. Wang, P . Perdikaris, Learning the solution operator of parametric partial differential equations with physics-informed deeponets, Sci. Adv. 7 (2021) eabi8605

    2021

  23. [23]

    Z. Li, H. Zheng, N. Kovachki, D. Jin, H. Chen, B. Liu, K. Azizzadenesheli, A. Anandkumar, Physics- informed neural operator for learning partial differential equations, ACM/IMS J. Data Sci. 1 (2024) 1–27

    2024

  24. [24]

    E. C. S. Acquesta, T. Portone, R. Dandekar, C. Rackauckas, R. Bandy, G. Huerta, I. Dytzel, Model- Form Epistemic Uncertainty Quantification for Modeling with Differential Equations: Application to Epidemiology, Technical Report, Sandia National Laboratories, 2022

    2022

  25. [25]

    Kaszás, U

    B. Kaszás, U. Feudel, T. Tél, Tipping phenomena in typical dynamical systems subjected to parameter drift, Sci. Rep. 9 (2019) 8654

    2019

  26. [26]

    M. C. Kennedy, A. O’Hagan, Bayesian calibration of computer models, J. R. Stat. Soc. Ser. B Stat. Methodol. 63 (2001) 425–464

    2001

  27. [27]

    Y. Ling, J. Mullins, S. Mahadevan, Selection of model discrepancy priors in bayesian calibration, J. Comput. Phys. 276 (2014) 665–680

    2014

  28. [28]

    M. R. Ebers, K. M. Steele, J. N. Kutz, Discrepancy modeling framework: Learning missing physics, modeling systematic residuals, and disambiguating between deterministic and random effects, SIAM J. Appl. Dyn. Syst. 23 (2024) 440–469

    2024

  29. [29]

    Boullé, D

    N. Boullé, D. Halikias, S. E. Otto, A. Townsend, Operator learning without the adjoint, J. Mach. Learn. Res. 25 (2024) 1–54

    2024

  30. [30]

    Boullé, D

    N. Boullé, D. Halikias, A. Townsend, Elliptic PDE learning is provably data-efficient, Proc. Natl. Acad. Sci. U. S. A. 120 (2023) e2303904120

    2023

  31. [31]

    Boullé, S

    N. Boullé, S. Kim, T. Shi, A. Townsend, Learning Green’s functions associated with time-dependent partial differential equations, J. Mach. Learn. Res. 23 (2022) 1–34

    2022

  32. [32]

    Boullé, A

    N. Boullé, A. Townsend, Learning elliptic partial differential equations with randomized linear algebra, Found. Comput. Math. 23 (2023) 709–739

    2023

  33. [33]

    S. H. Rudy, S. L. Brunton, J. L. Proctor, J. N. Kutz, Data-driven discovery of partial differential equations, Sci. Adv. 3 (2017) e1602614

    2017

  34. [34]

    Champion, B

    K. Champion, B. Lusch, J. N. Kutz, S. L. Brunton, Data-driven discovery of coordinates and governing equations, Proc. Natl. Acad. Sci. U. S. A. 116 (2019) 22445–22451

    2019

  35. [35]

    Podina, B

    L. Podina, B. Eastman, M. Kohandel, A PINN approach to symbolic differential operator discovery with sparse data, arXiv preprint arXiv:2212.04630 (2022)

    Pith/arXiv arXiv 2022

  36. [36]

    Z. Zou, X. Meng, G. E. Karniadakis, Correcting model misspecification in physics-informed neural networks (PINNs), J. Comput. Phys. 505 (2024) 112918. 19

    2024

  37. [37]

    D. P . Kingma, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

    Pith/arXiv arXiv 2014

  38. [38]

    X. Meng, L. Yang, Z. Mao, J. del Águila Ferrandis, G. E. Karniadakis, Learning functional priors and posteriors from data and physics, J. Comput. Phys. 457 (2022) 111073

    2022

  39. [39]

    J. M. Burgers, The nonlinear diffusion equation: asymptotic solutions and statistical problems, Springer Science & Business Media, 2013

    2013

  40. [40]

    G. B. Whitham, Linear and nonlinear waves, John Wiley & Sons, 2011

    2011

  41. [41]

    R. J. LeVeque, Finite volume methods for hyperbolic problems, Cambridge University Press, 2002

    2002

  42. [42]

    J. Bec, K. Khanin, Burgers turbulence, Phys. Rep. 447 (2007) 1–66

    2007

  43. [43]

    L. Wang, J. Mi, X. Meng, Z. Guo, A localized mass-conserving lattice boltzmann approach for non- newtonian fluid flows, Commun. Comput. Phys. 17 (2015) 908–924

    2015

  44. [44]

    R. W. Ogden, Non-linear elastic deformations, Courier Corporation, 1997

    1997

  45. [45]

    G. A. Holzapfel, Nonlinear solid mechanics: a continuum approach for engineering, Wiley, 2000

    2000

  46. [46]

    Bouziani, N

    N. Bouziani, N. Boullé, Structure-preserving operator learning, arXiv preprint arXiv:2410.01065 (2024)

    Pith/arXiv arXiv 2024

  47. [47]

    X. Ren, X. Lyu, Mixed form based physics-informed neural networks for performance evaluation of two-phase random materials, Eng. Appl. Artif. Intell. 127 (2024) 107250

    2024

  48. [48]

    H. P . Langtangen, G. K. Pedersen, Scaling of differential equations, Springer, 2016

    2016

  49. [49]

    D. W. Abueidda, S. Koric, E. Guleryuz, N. A. Sobh, Enhanced physics-informed neural networks for hyperelasticity, Int. J. Numer. Meth. Eng. 124 (2023) 1585–1601. 20

    2023

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.