Pith. sign in

REVIEW 3 major objections 4 minor 48 references

LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries

T0 review · 3 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that for linear homogeneous PDEs, a model built as a linear combination of Lie-symmetry-transformed seed solutions satisfies the PDE exactly for every parameter setting, so solving an IBVP reduces to matching in

desk verdict Nice method with a genuine exact-PDE construction; the reliability claim about loss-to-error is asserted, not proven, but the paper is honest about it and deserves referee time. read the letter →

arxiv 2510.25731 v2 pith:CIFO25MU submitted 2025-10-29 cs.LG cs.AIcs.NAmath.NAphysics.comp-ph

René P. Klausen , Ivan Timofeev , Jonas Naujoks , Johannes Frank , Thomas Wiegand , Sebastian Lapuschkin , Wojciech Samek This is my paper
classification cs.LGcs.AIcs.NAmath.NAphysics.comp-ph MSC 35A3068T07
keywords Liesymmetriesinitial-boundaryvalueproblemslinearhomogeneousPDEsphysics-informedmachinelearningvariableprojectiongreedybasisselectionerrorestimationPINNs
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

LieSolver introduces a way to solve initial-boundary value problems by building the physics directly into the model architecture. Instead of penalizing PDE residuals during training, the model is a linear combination of functions generated by applying Lie symmetry transformations to known seed solutions; for linear homogeneous PDEs every such function is an exact solution, so the whole model satisfies the PDE for any parameter values. This reduces learning to fitting the initial and boundary data, and the authors show the boundary loss then becomes a faithful measure of solution accuracy. For well-posed problems, they argue it supports rigorous error estimates. On 1D heat and wave benchmarks, LieSolver reaches IBC mean-squared errors near 1e-6 with only a few dozen terms and runs one to two orders of magnitude faster than physics-informed neural networks, while producing compact and interpretable symbolic decompositions.

What carries the argument

The load-bearing object is the Lie point symmetry group of the PDE, computed via the infinitesimal criterion and the exponential map. These symmetries act on any solution to generate one-parameter families of new exact solutions; applied to a simple seed solution (e.g., a constant), they produce parametrized base solutions. LieSolver superposes these base solutions linearly, eliminates the linear amplitudes with variable projection (least squares), and refines the symmetry parameters with bound-constrained nonlinear least squares. Since the PDE is linear and homogeneous, superposition preserves exactness, so the model is an exact solution for every parameter setting.

What would settle it

Solve a well-posed linear homogeneous IBVP, such as the heat equation on a finite interval, with an initial condition whose Fourier spectrum lies outside the achievable parameter ranges of the chosen base catalogue (e.g., a very narrow Gaussian or a high-frequency mode with bounds excluding that scale). If LieSolver's IBC MSE plateaus well above 1e-6 while a classical spectral or finite-difference solver matches the data, the expressivity assumption fails.

Watch

Extended reading notes

Core claim

The central discovery is that Lie symmetries give a generative mechanism for exact solutions: starting from a seed solution such as the constant function, applying parametrized symmetry transformations produces families of functions that each solve the PDE, and because the PDE is linear and homogeneous, any linear combination of them is still an exact solution. LieSolver parameterizes the model directly in this solution space, so the PDE is satisfied by construction and the only remaining task is to fit the initial and boundary conditions. The paper argues that this makes the boundary loss a direct measure of solution error, yields rigorous error bounds for well-posed IBVPs, and produces com

Load-bearing premise

The preselected catalogue of base solutions, together with its parameter bounds and sampling rules, must be expressive enough to represent the true solution; if the right shapes are missing, LieSolver cannot fit the data no matter how well the optimization runs.

Editorial extensions

If this is right

  • For any linear homogeneous PDE with a known symmetry group and a seed solution, LieSolver can produce models that satisfy the PDE exactly for all parameter values, eliminating the need for a PDE residual loss term.
  • The initial/boundary loss becomes a trustworthy training signal: a decrease in the loss directly implies an improvement in solution accuracy, and for well-posed IBVPs it bounds the domain-wide error rigorously.
  • Solutions are compact and interpretable: a few dozen symmetry-generated terms with their amplitudes form a sparse symbolic expression, in contrast to the opaque parameterizations of standard neural PDE solvers.
  • On the reported 1D heat and wave benchmarks, LieSolver attains IBC MSE near 1e-6 for smooth initial conditions and runs one to two orders of magnitude faster than physics-informed neural networks.
  • Sharp initial conditions, such as step profiles, require more basis terms but still reach domain errors around 1e-3, where the physics-informed neural network baselines plateau between 1e-2 and 1e-1.
  • The greedy addition of bases, combined with variable projection, provides a practical optimization scheme that is stable enough to reach high accuracy without a PDE residual term.

Reading between the lines

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

  • A consequence the paper leaves implicit: the exact-PDE property rests on linearity, so extending LieSolver to nonlinear PDEs would require a different combination rule, because a sum of symmetry-generated exact solutions would no longer be exact in general.
  • The boundary-loss-to-error estimate could be turned into an automated stopping rule with a certified tolerance, using the well-posedness constant of the IBVP to decide when the model is accurate enough.
  • Since Lie point symmetries can be computed algorithmically and seed solutions for linear homogeneous PDEs are trivial (zero or constants), the base catalogue could in principle be generated automatically; the remaining bottleneck is selecting parameter ranges and sampling rules that cover the relevant solution scales.
  • The same greedy variable-projection scheme applies to any dictionary of exact PDE solutions, not only symmetry-generated ones, so the framework offers a general template for PDE-constrained learning with exact physics.
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 / 4 minor

Summary. LieSolver is a machine-learning method for linear homogeneous IBVPs that enforces the PDE exactly by construction. Base functions are generated by applying Lie-symmetry transformations to seed solutions, and the final model is a linear combination of such base functions; hence it solves the PDE for any parameter choice. Training reduces to fitting initial and boundary data by an IBC loss, using greedy base selection, ridge least squares for amplitudes, and nonlinear least squares for symmetry parameters. The paper claims that the IBC loss directly bounds the domain error for well-posed IBVPs, that this gives rigorous error estimation, and that LieSolver outperforms vanilla PINNs in speed and accuracy on 1D heat and wave benchmarks. The exact-PDE construction is mathematically sound; the loss-to-error claim is asserted but not proved.

Significance. The exact-PDE architecture is a genuinely useful idea for linear homogeneous problems: it turns the PDE residual into a hard constraint, produces compact and interpretable symbolic forms, and the provided code makes the experiments reproducible. If the loss-to-error claim were proved, it would be a substantial reliability improvement over PINNs. The paper also includes a self-contained derivation of the heat-equation symmetry group (Appendix A) and is honest about the dictionary-expressivity limitation and the absence of rigorous bounds. As it stands, the advertised rigor is not supported, and the empirical comparison needs more statistical care.

major comments (3)
  1. [Abstract; §1; §3.1, Eq. (29); Example 2.1] The manuscript's central reliability claim—'the loss directly quantifies domain-wide error' and 'enables rigorous error estimation for well-posed IBVPs'—is not established. Eq. (29) is an empirical MSE over finitely many boundary collocation points. To conclude a domain-error bound, one needs (i) a quantitative well-posedness/stability inequality in the relevant norm and (ii) a finite-sample estimate controlling the gap between that empirical loss and the continuous boundary residual. Neither is provided. Example 2.1 gives a Lipschitz constant only for a special heat-equation comparison and does not address finite sampling or nonhomogeneous boundary perturbations. The paper itself states in §1: 'We do not prove any bounds rigorously.' This contradicts the abstract and conclusion. Please add a precise theorem with assumptions, or explicitly rescope the claims to not assert rigor.
  2. [§3.2; §4.3; §5] Expressivity rests entirely on the hand-chosen catalogue S and the parameter bounds/sampling rules. The paper acknowledges this in §3.2 ('The initial catalogue of bases S determines the expressivity...') and §5, but the consequence is not analyzed: there is no completeness result ensuring the dictionary covers the true solution for the considered class. The experiments are consistent with the base families being matched to the IC profiles—§4.3 notes LieSolver 'naturally fits IC profiles with similar sine and Gaussian patterns.' Greedy selection and NLLS cannot recover a component absent from S. Please state the representable class (e.g., finite Fourier-mode coverage for the benchmarks), add failure-mode analysis, or reposition the contribution as a heuristic dictionary method whose reliability must be checked per problem.
  3. [§4.1–4.3, Tables 1–2] The 'outperforms PINNs' conclusion is based on single representative runs, with no seeds, error bars, or statistical significance. The PINN baseline is a vanilla fully connected network; details such as Adam learning rate, L-BFGS stopping criteria, and weight initialization are not reported, which limits reproducibility of the timing/accuracy comparison. The claim should be narrowed or supported by multiple runs and a specified PINN variant.
minor comments (4)
  1. [Eq. (23) vs. Eq. (64)] The T6 symmetry in Example 2.2 is written with denominator 1+4ϑt, while the exponential-map computation in Appendix A.3, Eq. (64) gives 1−4ϑt. If these are the same one-parameter group up to a sign re-definition of ϑ, state this explicitly; otherwise one of the formulas needs correction.
  2. [§4.3, after Fig. 3] The sentence 'we confirm the considerations from section 2.2 that there is an upper bound for the ratio between MSE on the domain and the test MSE on the IBCs...' overinterprets a single experiment. Eq. (13) is an example-specific stability inequality, not a proved general bound; please rephrase.
  3. [Algorithm 1 and §4.1] The hyperparameter 'nfev_global' (Table 1/§4.1) does not appear in Algorithm 1; clarify how the number of trust-region-reflective steps is used in the global NLLS refinement.
  4. [§2.2, Example 2.1] The domain notation for the reverse heat equation is inconsistent: Eq. (9)–(10) write x ∈ [0, 2π) with boundary values u(0,t)=u(2π,t)=0; later benchmarks use (0,1). Please be consistent about whether boundaries are included and whether Dirichlet conditions are homogeneous.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the exact-PDE construction is self-contained; the unproven loss-to-error claim is a rigor gap, not a circular reduction.

full rationale

The central derivation is not circular. LieSolver's exact-PDE property follows from standard linearity and symmetry facts: base functions are Lie-symmetry transforms of seed solutions, hence exact solutions, and finite linear combinations of solutions of a linear homogeneous PDE are again solutions (Section 3.1, Eq. (28)). The parameters (a and theta) are optimized only against the IBC loss (Eqs. (29)-(32)); no quantity derived from the target solution is fed back as a 'prediction'. The hand-chosen base catalogue is explicitly acknowledged as an expressivity limitation ('performance of LieSolver heavily depends on the chosen set of base solutions', Section 5), not a fitted input disguised as output. The main caveat is the repeated claim that the boundary loss 'directly quantifies domain-wide error' and enables 'rigorous error estimation'; the paper itself states in Section 1: 'The path towards theoretical convergence guarantees. We do not prove any bounds rigorously.' This is an unsupported stability/sampling assertion, not a definitional or self-citational reduction, and should be treated as a correctness/reliability risk rather than circularity. The only self-citation ([25], a PINN resampling paper co-authored by J. Naujoks) appears in a peripheral related-work list and is not load-bearing. Overall, the derivation is self-contained with respect to its main claim, so the circularity burden is low.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

LieSolver introduces no new physical entities. Its load-bearing assumptions are the expressiveness of the hand-chosen base-solution dictionary, the well-posedness/stability of the IBVPs, and the validity of the Lie-symmetry transformations on the domain. The main unquantified design freedom is the base catalog and its parameter bounds.

free parameters (4)
  • Base solution catalog S (seed solutions and Lie-transformation sequences) = Heat: f1 (scaled/shifted sine), f2 (Gaussian blob), f3 (Gaussian-modulated sine); Wave: f1 (standing wave), f2 (Gaussian
    The paper states in Sec 3.2: 'The initial catalogue of bases S determines the expressivity of the constructed solution, and thus is crucial for performance.' This is the main design freedom and directly controls approximation power.
  • Parameter bounds and sampling rules for symmetry parameters θ = Not reported in the paper; described only as 'should cover relevant scales' (Sec 3.2).
    Each base function has parameter bounds and a uniform/log-uniform sampling rule that are design choices; the paper gives no concrete values, making this an unquantified free choice.
  • Hyperparameters: P, R, nfev_global, λ, L, mse_tol = P=1000, R=5, nfev_global=4, λ=1e-1, L=3000, mse_tol=1e-6 (Sec 4.1).
    These are empirically determined and affect the greedy search, refinement, and stopping behavior. They are not derived from theory.
  • PINN baseline IBC weight w_ICBC = 1e3 to 1e5, tuned per problem (Table 2).
    The baseline comparison depends on a tuned loss weight; the paper reports the tuned values, but this is still a fitted constant in the comparison.
assumptions (5)
  • standard math The infinitesimal criterion and Lie's theorems correctly characterize all continuous symmetries of the considered PDEs (Olver's theory).
    Used throughout Sec 2.3 and Appendix A.1 to justify that the transformations T1–T6 and T1–T2 are genuine symmetry groups.
  • domain assumption The studied IBVPs are well-posed and admit a stability estimate of the form ||u−v|| ≤ C||u(0)−v(0)||.
    The error-bound argument relies on continuous dependence on initial/boundary data. The paper demonstrates this for the heat equation example (Example 2.1) but does not prove it for all test cases, especially the wave equation.
  • domain assumption The Lie-symmetry-generated functions remain well-defined and smooth on the entire space-time domain for the chosen parameter bounds.
    Transformations like T6 contain denominators (1+4θt); the paper notes this in Example 2.2 and says bounds are restricted, but does not state the exact global conditions used in experiments.
  • domain assumption Uniform sampling on the boundary with L=3000 points sufficiently represents the continuous boundary conditions.
    The loss is evaluated on finite point sets Xk; the claim that boundary loss quantifies domain error requires these samples to be representative of the whole boundary.
  • domain assumption The Fourier-series reference solutions (Eqs. 33 and 34) with M terms are accurate enough to serve as ground truth for L2RE evaluation.
    The reference is computed by projecting initial data onto sine modes; truncation error is not reported, so the ground truth is itself approximate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries." pith.science (2026). https://pith.science/paper/CIFO25MU

@misc{pith2026251025731,
  author       = {Pith},
  title        = {Pith review of: LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CIFO25MU}},
  note         = {Machine review of arXiv:2510.25731}
}
read the original abstract

Initial-boundary value problems (IBVPs) provide the essential framework for modelling a wide range of phenomena in physics and engineering. We introduce a novel method for efficiently solving IBVPs using Lie symmetries to enforce the associated partial differential equation (PDE) exactly by construction. By leveraging symmetry transformations, our model embeds the underlying physical laws and learns the solution solely from initial and boundary data. Consequently, the boundary loss directly quantifies domain-wide error, enabling rigorous error estimation for well-posed IBVPs. We implement LieSolver and demonstrate its application to linear homogeneous PDEs, showing that it outperforms physics-informed neural networks (PINNs) in both speed and accuracy while yielding compact models. Overall, our approach significantly enhances the efficiency and reliability of predictions for PDE-constrained problems.

Figures

Figures reproduced from arXiv: 2510.25731 by the authors.

Figure 1
Figure 1. Initial conditions (solution shapes at t = 0) for the test cases for the heat (top row) and the wave (bottom row) equations. Code availability The Python implementation is available at https://github.com/oduwancheekee/liesolver. 4 Experiments and results To evaluate performance and demonstrate the capabilities of the model, we apply LieSolver to several IBVPs for the one-dimensional heat equation, and the one-dimens… view at source ↗
Figure 2
Figure 2. Representative base families for one-dimensional heat equation at [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Fit progress of LieSolver for the heat equation withGaussian IC: IBC MSE is shown as a function of the number of added bases. The vertical drops correspond to the global refinement steps [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: 2D domain fields (prediction, ground truth, and error) for the heat equation with [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: IBC fits for the heat equation with Gaussian IC, including the decomposition into weighted base contributions. The transparent blue curves depict the individual contributions. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the PINN prediction with the ground truth on the IBCs for the heat [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Training progress of LieSolver for the heat equation with step IC. IBC MSE is shown as a function of the number of added bases. Heat equation with step IC The step IC represents the most challenging case due to the jump in the profile [PITH_FULL_IMAGE:figures/full_fig…
Figure 8
Figure 8. Figure 8: 2D domain fields (prediction, ground truth, and error) for the heat equation with [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]
Figure 9
Figure 9. Figure 9: IBC fits for the heat equation with step IC, including the decomposition into weighted base contributions. The transparent blue curves depict the individual contributions [PITH_FULL_IMAGE:figures/full_fig_p018_9.png]
Figure 10
Figure 10. Figure 10: Comparison of the PINN prediction with the ground truth on the IBCs for the heat [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Training progress of LieSolver for the wave equation with step IC. IBC MSE is shown as a function of the number of added bases. Wave equation with step IC [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: 2D domain fields (prediction, ground truth, and error) for the wave equation with step [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]
Figure 13
Figure 13. Figure 13: IBC fits for the wave equation with step IC, including the decomposition into weighted [PITH_FULL_IMAGE:figures/full_fig_p020_13.png]
Figure 14
Figure 14. Figure 14: Comparison of the PINN prediction with the ground truth on the IBCs for the wave [PITH_FULL_IMAGE:figures/full_fig_p020_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 2 canonical work pages

  1. [1]

    Lie Point Symmetry and Physics-Informed Networks

    T. Akhound-Sadegh, L. Perreault-Levasseur, J. Brandstetter, M. Welling, and S. Ravan- bakhsh. “Lie Point Symmetry and Physics-Informed Networks”. In: Advances in Neural In- formation Processing Systems 36 (2023), pp. 42468–42481. arXiv: 2311.04293

    arXiv 2023

  2. [2]

    G. W. Bluman, A. F. Cheviakov, and S. C. Anco. Applications of symmetry methods to partial differential equations. Applied mathematical sciences. New York, NY: Springer, 2010

    2010

  3. [3]

    G. W. Bluman and S. Kumei. Symmetries and differential equations . Applied Mathematical Sciences. Berlin, Germany: Springer, 1989

    1989

  4. [4]

    W. E. Boyce. Elementary differential equations and boundary value problems, eleventh edition loose-leaf print companion. 11th ed. Nashville, TN: John Wiley & Sons, 2016

    2016

  5. [5]

    Lie Point Symmetry Data Augmentation for Neural PDE Solvers

    J. Brandstetter, M. Welling, and D. E. Worrall. “Lie Point Symmetry Data Augmentation for Neural PDE Solvers”. In: Proceedings of the 39th International Conference on Machine Learning. International Conference on Machine Learning. PMLR, 2022, pp. 2241–2256

    2022

  6. [6]

    A comparative study of some computer algebra packages which determine the Lie point symmetries of differential equations

    J. Butcher, J. Carminati, and K. Vu. “A comparative study of some computer algebra packages which determine the Lie point symmetries of differential equations”. In: Computer Physics Communications 155.2 (2003), pp. 92–114. doi: https : / / doi . org / 10 . 1016 / S0010 - 4655(03)00348-5

    2003

  7. [7]

    The computer calculation of Lie point symmetries of large systems of differential equations

    B. Champagne, W. Hereman, and P. Winternitz. “The computer calculation of Lie point symmetries of large systems of differential equations”. In: Computer Physics Communications 66.2 (1991), pp. 319–340. doi: https://doi.org/10.1016/0010-4655(91)90080-5

    work page doi:10.1016/0010-4655(91)90080-5 1991

  8. [8]

    Group equivariant convolutional networks

    T. Cohen and M. Welling. “Group equivariant convolutional networks”. In: International Conference on Machine Learning . PMLR. 2016, pp. 2990–2999

    2016

Show all 48 references
  1. [9]

    Lagrangian neural networks

    M. Cranmer, S. Greydanus, S. Hoyer, P. Battaglia, D. Spergel, and S. Ho. “Lagrangian neural networks”. In: International Conference on Learning Representations . 2020

    2020

  2. [10]

    J. J. Duistermaat and J. A. C. Kolk. Lie groups. Berlin, Germany: Springer, 1999

    1999

  3. [11]

    L. C. Evans. Partial Differential Equations . 2nd. Vol. 19. Graduate Studies in Mathematics. American Mathematical Society, 2010

    2010

  4. [12]

    The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate

    G. H. Golub and V. Pereyra. “The differentiation of pseudo-inverses and nonlinear least squares problems whose variables separate”. In: SIAM Journal on numerical analysis 10.2 (1973), pp. 413–432

    1973

  5. [13]

    Hamiltonian neural networks

    S. Greydanus, M. Dzamba, and D. Yosinski. “Hamiltonian neural networks”. In: Advances in Neural Information Processing Systems . Vol. 32. 2019

    2019

  6. [14]

    Symbolic Software for Lie Symmetry Analysis

    W. Hereman. “Symbolic Software for Lie Symmetry Analysis”. In: CRC Handbook of Lie Group Analysis of Differential Equations . Ed. by N. Ibragimov. Vol. III. CRC Press, 1996, pp. 367–413

    1996

  7. [15]

    Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations

    A. D. Jagtap, E. Kharazmi, and G. E. Karniadakis. “Extended physics-informed neural networks (XPINNs): A generalized space-time domain decomposition based deep learning framework for nonlinear partial differential equations”. In:Communications in Computational Physics 28.5 (20...

    2020

  8. [16]

    F. John. Partial Differential Equations. Applied mathematical sciences. New York, NY: Springer US, 1978

    1978

  9. [17]

    Physics- informed machine learning

    G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang. “Physics- informed machine learning”. In: Nature Reviews Physics 3.6 (2021), pp. 422–440. doi: 10. 1038/s42254-021-00314-5

    2021

  10. [18]

    Characterizing possible failure modes in physics-informed neural networks

    A. S. Krishnapriyan, A. Gholami, S. Zhe, R. M. Kirby, and M. W. Mahoney. “Characterizing possible failure modes in physics-informed neural networks”. In: (2021). arXiv: 2109.01050. 28

    2021 arXiv

  11. [19]

    Artificial Neural Networks for Solving Ordinary and Partial Differential Equations

    I. E. Lagaris, A. Likas, and D. I. Fotiadis. “Artificial Neural Networks for Solving Ordinary and Partial Differential Equations”. In: IEEE Transactions on Neural Networks 9.5 (1998), pp. 987–1000. doi: 10.1109/72.712178. arXiv: physics/9705023

    1998 arXiv

  12. [20]

    PINNACLE: PINN Adaptive ColLocation and Experimental Points Selection

    G. K. R. Lau, A. Hemachandra, S.-K. Ng, and B. K. H. Low. “PINNACLE: PINN Adaptive ColLocation and Experimental Points Selection”. In: The Twelfth International Conference on Learning Representations. 2023

    2023

  13. [21]

    R. J. LeVeque. Finite Difference Methods for Ordinary and Partial Differential Equations: Steady-State and Time-Dependent Problems . Philadelphia, PA: Society for Industrial and Applied Mathematics, 2007. doi: 10.1137/1.9780898717839

    2007 doi

  14. [22]

    Physics-informed neural networks with hard constraints for inverse design

    L. Lu, R. Pestourie, W. Yao, Z. Wang, F. Verdugo, and S. G. Johnson. “Physics-informed neural networks with hard constraints for inverse design”. In: (2021). arXiv: 2102.04626

    2021 arXiv

  15. [23]

    PPINNs: Physics-informed neural networks with prior

    X. Meng, Z. Li, and G. E. Karniadakis. “PPINNs: Physics-informed neural networks with prior”. In: Journal of Computational Physics 401 (2020), p. 109020

    2020

  16. [24]

    Self-Supervised Learning with Lie Symmetries for Partial Differential Equations

    G. Mialon, Q. Garrido, H. Lawrence, D. Rehman, Y. LeCun, and B. Kiani. “Self-Supervised Learning with Lie Symmetries for Partial Differential Equations”. In: Advances in Neural Information Processing Systems 36 (2023), pp. 28973–29004

    2023

  17. [25]

    J. R. Naujoks et al. Leveraging Influence Functions for Resampling Data in Physics-Informed Neural Networks . 2025. doi: 10 . 48550 / arXiv . 2506 . 16443. arXiv: cs/2506 . 16443. Pre- published

    2025

  18. [26]

    Variable projection for nonlinear least squares problems

    D. P. O’Leary and B. W. Rust. “Variable projection for nonlinear least squares problems”. In: Computational Optimization and Applications 54.3 (2013), pp. 579–593

    2013

  19. [27]

    Lie Symmetries of Differential Equations: Classical Results and Recent Contri- butions

    F. Oliveri. “Lie Symmetries of Differential Equations: Classical Results and Recent Contri- butions”. In: Symmetry 2.2 (2010), pp. 658–706. doi: 10.3390/sym2020658

    2010 doi

  20. [28]

    ReLie: a Reduce program for Lie group analysis of differential equations

    F. Oliveri. “ReLie: a Reduce program for Lie group analysis of differential equations”. In: (2021). arXiv: 2105.11534

    2021 arXiv

  21. [29]

    P. J. Olver. Introduction to partial differential equations . en. 1st ed. Undergraduate texts in mathematics. Cham, Switzerland: Springer International Publishing, 2013

    2013

  22. [30]

    Symmetry and explicit solutions of partial differential equations

    P. J. Olver. “Symmetry and explicit solutions of partial differential equations”. In: Applied Numerical Mathematics 10.3 (1992), pp. 307–324. doi: https://doi.org/10.1016/0168- 9274(92)90047-H

    1992 doi

  23. [31]

    P. J. Olver. Applications of Lie Groups to Differential Equations . 2nd. New York: Springer- Verlag, 1993

    1993

  24. [32]

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

    M. Raissi, P. Perdikaris, and G. E. Karniadakis. “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial dif- ferential equations”. In: Journal of Computational Physics 378 (2019), pp. 686–707. doi: 10.1...

    2019 doi

  25. [33]

    Challenges in Training PINNs: A Loss Landscape Perspective

    P. Rathore, W. Lei, Z. Frangella, L. Lu, and M. Udell. “Challenges in Training PINNs: A Loss Landscape Perspective”. In: Proceedings of the 41st International Conference on Machine Learning. Ed. by R. Salakhutdinov, Z. Kolter, K. Heller, A. Weller, N. Oliver, J. Scarlett, and ...

    2024

  26. [34]

    J. Rauch. Partial Differential Equations. en. Graduate Texts in Mathematics. New York, NY: Springer, 1991

    1991

  27. [35]

    DGM: A deep learning algorithm for solving partial dif- ferential equations

    J. Sirignano and K. Spiliopoulos. “DGM: A deep learning algorithm for solving partial dif- ferential equations”. In: Journal of Computational Physics 375 (2018), pp. 1339–1364. doi: 10.1016/j.jcp.2018.08.029. eprint: 1708.07469. 29

    2018 arXiv

  28. [36]

    Practical Guide to the Symbolic Computation of Symmetries of Differential Equations

    S. Steinberg and R. de Melo Marinho Junior. “Practical Guide to the Symbolic Computation of Symmetries of Differential Equations”. In: (2014). arXiv: 1409.8364

    2014 arXiv

  29. [37]

    Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds

    N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley. “Tensor field networks: Rotation- and translation-equivariant neural networks for 3D point clouds”. In: Advances in Neural Information Processing Systems . Vol. 31. 2018

    2018

  30. [38]

    Solving Differential Equations with Genetic Programming

    I. G. Tsoulos and I. E. Lagaris. “Solving Differential Equations with Genetic Programming”. In: Genetic Programming and Evolvable Machines 7.1 (2006), pp. 33–54. doi: 10 . 1007 / s10710-006-7009-y

    2006

  31. [39]

    Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective

    S. Wang, A. K. Bhartari, B. Li, and P. Perdikaris. “Gradient Alignment in Physics-informed Neural Networks: A Second-Order Optimization Perspective”. In: (2025). arXiv: 2502.00604

    2025

  32. [40]

    An Expert’s Guide to Training Physics- informed Neural Networks

    S. Wang, S. Sankaran, H. Wang, and P. Perdikaris. “An Expert’s Guide to Training Physics- informed Neural Networks”. In: (2023). arXiv: 2308.08468

    2023 arXiv

  33. [41]

    Understanding and Mitigating Gradient Flow Patholo- gies in Physics-Informed Neural Networks

    S. Wang, Y. Teng, and P. Perdikaris. “Understanding and Mitigating Gradient Flow Patholo- gies in Physics-Informed Neural Networks”. In: SIAM Journal on Scientific Computing 43.5 (2021), A3055–A3081. doi: 10.1137/20M1318043

    2021 doi

  34. [42]

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

    S. Wang, Y. Teng, and P. Perdikaris. “When and why PINNs fail to train: A neural tangent kernel perspective”. In: Journal of Computational Physics 449 (2022), p. 110768

    2022

  35. [43]

    E. P. Wigner. Symmetries and reflections: scientific essays . Indiana University Press, 1967

    1967

  36. [44]

    A Comprehensive Study of Non-Adaptive and Residual-Based Adaptive Sampling for Physics-Informed Neural Networks

    C. Wu, M. Zhu, Q. Tan, Y. Kartha, and L. Lu. “A Comprehensive Study of Non-Adaptive and Residual-Based Adaptive Sampling for Physics-Informed Neural Networks”. In:Computer Methods in Applied Mechanics and Engineering 403 (2023), p. 115671. doi: 10.1016/j.cma. 2022.115671

    2023

  37. [45]

    Latent Space Symmetry Discovery

    J. Yang, N. Dehmamy, R. Walters, and R. Yu. “Latent Space Symmetry Discovery”. In: Proceedings of the 41st International Conference on Machine Learning . Vol. 235. ICML’24. Vienna, Austria: JMLR.org, 2024, pp. 56047–56070

    2024

  38. [46]

    Symmetry-Informed Governing Equa- tion Discovery

    J. Yang, W. Rao, N. Dehmamy, R. Walters, and R. Yu. “Symmetry-Informed Governing Equa- tion Discovery”. In: Advances in Neural Information Processing Systems 37 (2024), pp. 65297– 65327

    2024

  39. [47]

    A Symmetry Group Based Supervised Learning Method for Solving Partial Differential Equations

    Z.-Y. Zhang, S.-J. Cai, and H. Zhang. “A Symmetry Group Based Supervised Learning Method for Solving Partial Differential Equations”. In: Computer Methods in Applied Me- chanics and Engineering 414 (2023), p. 116181. doi: 10.1016/j.cma.2023.116181

    2023

  40. [48]

    Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equations

    Z.-Y. Zhang, H. Zhang, L.-S. Zhang, and L.-L. Guo. “Enforcing continuous symmetries in physics-informed neural network for solving forward and inverse problems of partial differential equations”. In: Journal of Computational Physics 492 (2023), p. 124145. 30

    2023

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.