Pith. sign in

REVIEW 4 major objections 6 minor 41 references

IP-Basis PINNs: Efficient Multi-Query Inverse Parameter Estimation

T0 review · 4 major / 6 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 claims that a single offline-trained network can act as a reusable basis so that each new inverse problem—recovering unknown parameters of a differential equation from data—is solved by training only a thin linear layer, giving a

desk verdict A sensible inverse-problem extension of Desai et al.'s one-shot transfer learning, with real engineering value and honest limitations, but the evidence is thin and the core 'basis' assumption is never actually tested. read the letter →

arxiv 2509.07245 v1 pith:KXZJVYL3 submitted 2025-09-08 cs.LG

Shalev Manor , Mohammad Kohandel This is my paper
classification cs.LG MSC 65M3268T07
keywords physics-informedneuralnetworksinverseproblemsparameterestimationmulti-queryoffline-onlinedecompositionbasisfunctionsforward-modeautomaticdifferentiationmeta-learning
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

The paper claims that inverse parameter estimation—recovering unknown coefficients or functional terms in a differential equation from data—can be made cheap enough for repeated use by splitting the work into two phases. Offline, one network is trained so that several of its outputs act as a basis for the solution family of a parametric equation. Online, that network is frozen, and for each new dataset only a thin linear output layer and the unknown parameter estimates are trained, with the PDE residual evaluated through precomputed derivatives. On a damped oscillator, a Lotka-Volterra system with unknown interaction terms, and the quantum harmonic oscillator, the method recovers parameters accurately and reports per-query training times about an order of magnitude lower than a standard PINN (about 22x in one comparison), while tolerating sparse and noisy data. The paper is explicit that the 'basis' property is not enforced, only learned, so the method's validity rests on the offline network actually spanning the solution space for the parameter range encountered online.

What carries the argument

The load-bearing object is the pair (R, L): a shared feature network R followed by a linear readout layer L. Offline, multiple readouts are trained on multiple parameter samples of the same differential equation, so the loss is the empirical average of the infimum over linear maps of the PINN loss for each parameter; this shapes R's outputs into a spanning set for the solution family. Online, R is frozen; its values and derivatives at all collocation points are precomputed once, and only a new readout layer and the parameter estimates are optimized. The online loss combines a data term with a residual term evaluated at the current parameter estimates, and forward-mode (hyper-dual) automatic

What would settle it

Fix a parameter value well outside the offline training range (for example, damping coefficients above 1.5 in the oscillator test or force constants outside 0-5 in the quantum oscillator), run the online procedure with abundant clean data, and check whether both the recovered parameters and the reconstructed solution stay near their in-distribution error levels. A more direct test is algebraic: compute the Gram matrix of R's outputs (or of its Jacobian) at the collocation points; if its condition number is large or its numerical rank is well below the number of readouts, the linear feature map

Watch

Extended reading notes

Core claim

The central claim is that a frozen network, whose many outputs were trained offline to satisfy the same parametric PDE for many parameter samples, can be reused as a linear basis for solving new inverse problems. Given a new dataset, the method keeps R fixed and optimizes only a new readout layer plus scalar parameter estimates, using a data-fit term and a PDE residual that depends on those parameter estimates. Because R's outputs and their derivatives at collocation points are computed and stored once, each online gradient step costs a few matrix multiplications rather than a full backpropagation pass; forward-mode automatic differentiation lets all output derivatives be obtained in a singl

Load-bearing premise

The method works only if the offline-trained network's output features span the solution family of the parametric PDE for the parameter values actually seen online; the paper does not enforce or verify this spanning property, and its out-of-distribution trial shows accuracy falling sharply when the assumption is violated.

Editorial extensions

If this is right

  • After one offline training run, each new dataset for the same parametric equation family requires training only a linear layer and a few scalar parameters, removing the full-network retraining cost of standard PINNs.
  • The paper's quantum-oscillator comparison indicates per-query training time can drop by roughly a factor of 22, with parameter error comparable to standard PINNs at 1,000 or more data points.
  • The method extends beyond constant coefficients: with universal-PINN networks substituted for unknown interaction terms, Lotka-Volterra parameters are recovered along with the functional terms themselves.
  • Sparse data are usable—accuracy degrades gracefully from 10,000 down to 100 points in the oscillator test, and with 10 points the basis method avoids the catastrophic failure a standard PINN showed.
  • Noisy data degrade parameter accuracy smoothly, making the approach suitable for measurements with sensor noise; the paper shows this for the quantum harmonic oscillator.

Reading between the lines

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

  • Because the online phase is just a linear fit on frozen features, the readout weights could likely be obtained by closed-form least squares when the PDE is linear, removing online gradient descent altogether; the paper notes the analogous closed form exists for the forward problem but does not extend it to inverse problems.
  • The sharp drop in out-of-distribution accuracy suggests that a practical deployment should include a validity check—e.g., the conditioning of the feature Gram matrix at collocation points—before trusting an online estimate.
  • The same precomputation-plus-linear-readout recipe could be applied to a randomized, untrained feature map; comparing that baseline against the physics-trained basis would isolate how much of the speedup and robustness comes from the offline physics training rather than from the linear readout structure.
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 paper proposes IP-Basis PINNs, an offline-online meta-learning framework for repeated solution of inverse parameter estimation problems for parametric ODEs/PDEs. Offline, a shared feature network R is trained with multiple linear readouts to act as a basis for the solution manifold of a parametric differential equation. Online, R is frozen; its outputs and derivatives are precomputed, and for each new data set only a new linear readout layer and the unknown PDE parameters are trained, using a data loss plus a PDE residual loss evaluated with the estimated parameters. The paper claims three main contributions: a new online loss for simultaneous solution reconstruction and parameter identification, the use of forward-mode automatic differentiation for efficient Hessian/PDE-loss evaluation, and a validation-based early-stopping mechanism for offline training. Experiments are reported on a damped harmonic oscillator, a Lotka-Volterra system with unknown interaction terms via Universal PINNs, and the quantum harmonic oscillator, with claims of significant per-query speedup and robustness to scarce and noisy data.

Significance. If the central claim holds, the method is a useful extension of one-shot transfer-learning PINNs to inverse problems: a single offline investment yields a reusable feature representation, and online inference reduces to optimizing a linear layer and a few scalar parameters, which is attractive for many-query and real-time settings. The forward-mode AD speedup and the validation heuristic are sensible contributions, and the paper explicitly acknowledges limitations of the learned 'basis' and the lack of an empirical comparison with GPT-PINN. However, the current evidence is thin: no code or data are released, the main quantitative claims rest on single runs, and the load-bearing assumption that the offline-trained R spans the solution space for unseen parameters is not directly verified.

major comments (4)
  1. [Section 2.1 and Section 4] The entire online phase depends on the assumption that, for every parameter p in the target distribution, there exists a linear readout T such that T∘R approximates u(p) and satisfies the PDE at p. Section 2.1 merely states that R is 'treated as a basis' and optimizes an empirical average over n sampled parameters; it does not test whether the span of R contains good approximations for held-out p. Section 4 admits that linear independence is not explicitly enforced and that the outputs may be 'a more general spanning set.' This matters not just terminologically: if the span misses directions needed for an unseen p, online optimization cannot recover the solution or the parameter. The paper should provide a direct diagnostic, e.g., report inf_T L_PINN(p, T∘R) on a fresh set of parameters, the representation error of held-out solutions, or the residual norm as a function of the number of r
  2. [Section 2.4] The validation loss ilde{L} is defined as an average of L_PINN( ilde{p}_i, ( ilde{L}∘R)), but the quantity the method really wants is inf_T L_PINN( ilde{p}, T∘R). In practice ilde{L} is optimized by only one gradient step per offline epoch, starting from a random initialization. A one-step-trained linear layer is not generally close to the infimum over T, and no evidence is provided that this one-step loss tracks the true infimum. Since early stopping is listed as a contribution and used to select the final offline model, the approximation should be validated, for example by periodically comparing the one-step validation score with a fully converged ilde{L} on a small subset of validation parameters.
  3. [Section 3, Tables 1-7] All experimental numbers are reported without error bars or repeated-seed statistics. For a stochastic training process, and given the paper's own observation in Section 3.3 that 'the choice of weights seemed to affect the final outcome of training significantly,' single trials are insufficient to support the abstract's claims of 'consistent performance' and 'robust operation.' This is particularly problematic in Table 2, where R10 performs best in-distribution but R30 and R50 degrade more gracefully out-of-distribution; without variance estimates these patterns may be noise. Please report mean and standard deviation over multiple independent offline and online runs, or at least over multiple data realizations and random seeds.
  4. [Section 3.3, speedup comparison] The claimed ~22x inference speedup is computed by taking the 187-second online time for 10 IP-Basis readouts and comparing it with a single standard-PINN run of 421 seconds, extrapolated to 10 runs. The standard PINN hyperparameters were tuned only for k=5, and only one test parameter (k=2.666...) was used for the comparison. In addition, the comparison mixes epochs (6000 for IP-Basis online vs 5000 for the standard PINN) and does not specify whether the same collocation points, loss weights, and optimizer settings were used. The per-query speedup is a central claim, so a fairer benchmark is needed: same hardware, same number of queries, several test parameters, and a statement of whether the offline amortization is included or excluded.
minor comments (6)
  1. [Section 3.2, after Table 5] The text says 'We plot 10 readouts trained online using R100 ... in Figure 1,' but the Lotka-Volterra results are plotted in Figure 2. Please correct the reference.
  2. [Table 3] The footnote marker 'R1001' is unclear; it should be typeset as a footnote reference to the backward-mode run, e.g., 'R100 (backward mode)'.
  3. [Acknowledgments] The name 'Micheal Alexander' should be 'Michael Alexander.'
  4. [Section 2.4] The phrase 'the network (L∘R) takes a step to optimize the training loss and ilde{L} takes a step to optimize the validation loss' is awkward: it is the combined network parameters that are updated. Please rephrase for clarity.
  5. [Section 3.3, Figure 4] The noise-robustness figure would benefit from error bars and from labeling the curves with the number of data points directly in the legend or caption; currently the reader must infer this from the text.
  6. [General] No code or data are released, and no random seeds are reported. While not a technical flaw per se, for an empirically driven paper this makes the quantitative claims difficult to reproduce or compare against future work.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: online parameter estimates come from a genuine inverse optimization, and the only self-citation (UPINN) is not load-bearing.

full rationale

The derivation chain is not circular. The offline phase optimizes R and L on an empirical average of PINN losses over sampled parameters p_i; this is a legitimate representation-learning prior, not a fit to the online query data. The online phase solves a standard inverse problem: it minimizes a data loss against observed measurements plus a PDE residual in which the unknown parameter estimate p̄ is a trainable variable. The parameter estimates are therefore outputs of an optimization, not quantities already contained in the offline training. No equation defines the predicted parameter in terms of the training loss; the offline/online decomposition is analogous to reduced-basis methods and is not self-referential. The only self-citation is [11] (UPINN), by co-author Kohandel, used to parameterize unknown Lotka-Volterra terms. This is not load-bearing circularity: the UPINN component is an externally published method, is described in the text, and is used only as a building block in one benchmark; no uniqueness or correctness claim for IP-Basis is derived from it. Section 4's admission that linear independence is not enforced is an honest limitation about the basis assumption, not a step that reduces the prediction to its inputs. The out-of-distribution degradation in Table 5 is empirical evidence that online inference is not trivially forced by offline training. Thus no circular step can be exhibited.

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

No new physical or mathematical entities are introduced. The learned basis functions are internal network representations, not external postulates. The free parameters are the usual hyperparameters and loss weights, selected by hand or on a small validation set for each experiment. The key domain assumption is that a finite set of readouts can form a basis for the solution family, which the paper itself acknowledges is not enforced or verified.

free parameters (5)
  • Loss weights (ω_PDE, ω_IC, ω_BC, ω_Data) = Various: ω_Data=1, ω_ODE=0.001 for harmonic oscillator; ω_PDE=0.1, ω_Data=1 for Lotka-Volterra; for QHO ω_PDE ramps from
    Chosen per experiment by hand/validation. In QHO (Section 3.3), the text says the choice of weights significantly affected the outcome and values were chosen to avoid trivial solutions.
  • Network architecture (width, depth, activation) = 4 hidden layers width 40 (harmonic), 4x64 (Lotka-Volterra), 5x100 (QHO), all tanh.
    Hyperparameters selected by hand; no architecture search or sensitivity analysis is reported.
  • Learning rates and training epochs = Offline: 5e-5 for 30k epochs (harmonic), 3e-4 for 40k (LV), 2e-3 with a schedule for 3.2k (QHO). Online: various schedul
    Tuned on a small validation set; no sensitivity analysis provided.
  • Offline parameter distribution ranges = α,β ∈ [0,1.5], f ∈ [-1.5,1.5] (harmonic); α,β,γ,δ ∈ [0.5,1.5] (LV); k ∈ {0,...,5} (QHO).
    Chosen to represent expected online inputs. Out-of-distribution tests (Table 5) show significant degradation, confirming the dependence on this choice.
  • Number of readouts in offline training = 10, 30, 50 for harmonic; 10, 30, 50, 100 for LV; 6 for QHO.
    Hyperparameter; the paper notes that more readouts do not consistently improve online accuracy (contrast Tables 1 and 2).
assumptions (6)
  • domain assumption A finite set of readouts trained on a sampled parameter distribution forms a basis for the solution space over that distribution.
    Invoked in Section 2.1 where R is treated as a basis for the solution space. Not proven or verified; Section 4 admits linear independence is not enforced and outputs may be only a spanning set.
  • ad hoc to paper The final linear layer L can be optimized to its global minimum, so the offline loss equals the empirical mean of inf_T LPINN(p_i, T∘R).
    Stated in Section 2.1 ('if the assumption is made that the final layer is optimized to its global minimum'). In practice, L is trained jointly with R, so the equality is only approximate.
  • ad hoc to paper The validation loss with one gradient step per epoch on a new readout L̃ approximates the true inf over T for early stopping.
    Section 2.4 acknowledges L̃ is not fully trained because doing so would increase training time; this one-step approximation is not validated.
  • standard math Hyper-dual numbers / forward-mode Taylor expansion gives correct first and second derivatives of the network output.
    Used in Section 2.3 to compute PDE residuals; based on standard Taylor series and the Fomoh library [10].
  • domain assumption The inverse problem is identifiable from the chosen data and residual losses.
    Implicit in the online phase (Section 2.2): minimizing L_data + L_PDE with trainable parameter estimates is assumed to recover true parameters; no identifiability analysis is given.
  • domain assumption PDE solutions are smooth enough for tanh networks and sampled collocation points to represent them accurately.
    The experiments use tanh activations and 10,000 collocation points for the QHO; no error analysis on discretization or function class is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of IP-Basis PINNs: Efficient Multi-Query Inverse Parameter Estimation." pith.science (2026). https://pith.science/paper/KXZJVYL3

@misc{pith2026250907245,
  author       = {Pith},
  title        = {Pith review of: IP-Basis PINNs: Efficient Multi-Query Inverse Parameter Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KXZJVYL3}},
  note         = {Machine review of arXiv:2509.07245}
}
read the original abstract

Solving inverse problems with Physics-Informed Neural Networks (PINNs) is computationally expensive for multi-query scenarios, as each new set of observed data requires a new, expensive training procedure. We present Inverse-Parameter Basis PINNs (IP-Basis PINNs), a meta-learning framework that extends the foundational work of Desai et al. (2022) to enable rapid and efficient inference for inverse problems. Our method employs an offline-online decomposition: a deep network is first trained offline to produce a rich set of basis functions that span the solution space of a parametric differential equation. For each new inverse problem online, this network is frozen, and solutions and parameters are inferred by training only a lightweight linear output layer against observed data. Key innovations that make our approach effective for inverse problems include: (1) a novel online loss formulation for simultaneous solution reconstruction and parameter identification, (2) a significant reduction in computational overhead via forward-mode automatic differentiation for PDE loss evaluation, and (3) a non-trivial validation and early-stopping mechanism for robust offline training. We demonstrate the efficacy of IP-Basis PINNs on three diverse benchmarks, including an extension to universal PINNs for unknown functional terms-showing consistent performance across constant and functional parameter estimation, a significant speedup per query over standard PINNs, and robust operation with scarce and noisy data.

Figures

Figures reproduced from arXiv: 2509.07245 by the authors.

Figure 1
Figure 1. Comparison between numerical solutions (dashed lines) and online model approximations (solid [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Comparison between numerical solutions (dashed lines) and online model approximations (solid [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Comparison between numerical solution (top row) and model predictions trained online for the [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Mean Absolute Error of predicted force constant plotted against the scale of injected noise. [PITH_FULL_IMAGE:figures/full_fig_p013_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages

  1. [1]

    GPT-PINN: Generative Pre-Trained Physics-Informed Neural Net- works toward non-intrusive Meta-learning of parametric PDEs

    Yanlai Chen and Shawn Koohy. “GPT-PINN: Generative Pre-Trained Physics-Informed Neural Net- works toward non-intrusive Meta-learning of parametric PDEs”. In:Finite Elements in Analysis and Design228 (Jan. 2024), p. 104047.issn: 0168874X.doi:10 . 1016 / j . finel . 2023 . 104047.url: https://linkinghub.elsevier.com/retrieve/pii/S0168874X23001403(visited on...

    work page 2024

  2. [2]

    Advanced Physics-informed neural networks for numerical approximation of the coupled Schr¨ odinger–KdV equation

    Qiongni Zhang et al. “Advanced Physics-informed neural networks for numerical approximation of the coupled Schr¨ odinger–KdV equation”. In:Communications in Nonlinear Science and Numerical Simulation138 (2024), p. 108229.issn: 1007-5704.doi:https://doi.org/10.1016/j.cnsns.2024. 108229.url:https://www.sciencedirect.com/science/article/pii/S1007570424004143

    doi:10.1016/j.cnsns.2024 2024

  3. [3]

    Fiber transmission model with parameterized inputs based on generative pre-trained physics-informed neural networks

    Yubin Zang et al. “Fiber transmission model with parameterized inputs based on generative pre-trained physics-informed neural networks”. In:Optics Express33.1 (Jan. 13, 2025), p. 50.issn: 1094-4087.doi: 10.1364/OE.539845.url:https://opg.optica.org/abstract.cfm?URI=oe-33-1-50(visited on 08/22/2025)

    work page doi:10.1364/oe.539845.url:https://opg.optica.org/abstract.cfm 2025

  4. [4]

    June 23, 2020.doi:10.48550/ arXiv.1911.02685

    Fuzhen Zhuang et al.A Comprehensive Survey on Transfer Learning. June 23, 2020.doi:10.48550/ arXiv.1911.02685. arXiv:1911.02685[cs].url:http://arxiv.org/abs/1911.02685(visited on 08/15/2025)

    arXiv doi:10.48550/arxiv.1911.02685 2020

  5. [5]

    July 5, 2022.doi: 10.48550/arXiv.2110.11286

    Shaan Desai et al.One-Shot Transfer Learning of Physics-Informed Neural Networks. July 5, 2022.doi: 10.48550/arXiv.2110.11286. arXiv:2110.11286[cs].url:http://arxiv.org/abs/2110.11286 (visited on 05/27/2025)

    arXiv doi:10.48550/arxiv.2110.11286 2022

  6. [6]

    Meta-learning Loss Functions of Parametric Partial Differential Equations Using Physics-Informed Neural Networks

    Michail Koumpanakis and Ricardo Vilalta. “Meta-learning Loss Functions of Parametric Partial Dif- ferential Equations Using Physics-Informed Neural Networks”. In:ArXivabs/2412.00225 (2024).url: https://api.semanticscholar.org/CorpusID:274437009

    work page Pith review arXiv 2024

  7. [7]

    Automatic Differentiation in Machine Learning: a Survey

    Atilim Gunes Baydin et al. “Automatic Differentiation in Machine Learning: a Survey”. In:Journal of Machine Learning Research18.153 (2018), pp. 1–43.issn: 1533-7928.url:http://jmlr.org/papers/ v18/17-468.html(visited on 07/29/2025)

    work page 2018

  8. [8]

    PyTorch: An Imperative Style, High-Performance Deep Learning Library

    Adam Paszke et al. “PyTorch: An Imperative Style, High-Performance Deep Learning Library”. In: Advances in Neural Information Processing Systems. Ed. by H. Wallach et al. Vol. 32. Curran As- sociates, Inc., 2019.url:https : / / proceedings . neurips . cc / paper _ files / paper / 2019 / file / bdbca288fee7f92f2bfa9f7012727740-Paper.pdf

    work page 2019

Show all 41 references
  1. [9]

    Software available from tensorflow.org

    Mart ´ ın Abadi et al.TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. Software available from tensorflow.org. 2015.url:https://www.tensorflow.org/

    2015

  2. [10]

    Cobb et al.Second-Order Forward-Mode Automatic Differentiation for Optimization

    Adam D. Cobb et al.Second-Order Forward-Mode Automatic Differentiation for Optimization. Aug. 19, 2024.doi:10.48550/arXiv.2408.10419. arXiv:2408.10419[cs].url:http://arxiv.org/abs/ 2408.10419(visited on 07/28/2025)

    2024 doi arXiv

  3. [11]

    Universal Physics-Informed Neural Net- works: Symbolic Differential Operator Discovery with Sparse Data

    Lena Podina, Brydon Eastman, and Mohammad Kohandel. “Universal Physics-Informed Neural Net- works: Symbolic Differential Operator Discovery with Sparse Data”. In:Proceedings of the 40th In- ternational Conference on Machine Learning. International Conference on Machine Learnin...

    2023

  4. [12]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential 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 differential equations”. In:Journal of Computational Physics378 (Feb. 2019), pp. 686–707.issn: 0021...

    2019 doi

  5. [13]

    Baratta et al.DOLFINx: the next generation FEniCS problem solving environment

    Igor A. Baratta et al.DOLFINx: the next generation FEniCS problem solving environment. preprint. 2023.doi:10.5281/zenodo.10447666

    2023 doi

  6. [14]

    A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems

    Peter Benner, Serkan Gugercin, and Karen Willcox. “A Survey of Projection-Based Model Reduction Methods for Parametric Dynamical Systems”. In:SIAM Review57.4 (Jan. 2015). Publisher: Society for Industrial and Applied Mathematics, pp. 483–531.issn: 0036-1445.doi:10.1137/1309327...

    2015 doi

  7. [15]

    The”echo state

    Herbert Jaeger. “The”echo state”approach to analysing and training recurrent neural networks”. In: 2001.url:https://api.semanticscholar.org/CorpusID:15467150. 16

    2001

  8. [16]

    Next generation reservoir computing

    Daniel J. Gauthier et al. “Next generation reservoir computing”. In:Nature Communications12.1 (Sept. 21, 2021). Publisher: Springer Science and Business Media LLC.issn: 2041-1723.doi:10 . 1038 / s41467 - 021 - 25801 - 2.url:https : / / www . nature . com / articles / s41467 - ...

    2021

  9. [17]

    Task-adaptive physical reservoir computing

    Oscar Lee et al. “Task-adaptive physical reservoir computing”. In:Nature Materials23.1 (Jan. 2024). Publisher: Springer Science and Business Media LLC, pp. 79–87.issn: 1476-1122, 1476-4660.doi: 10.1038/s41563-023-01698-8.url:https://www.nature.com/articles/s41563-023-01698-8 (...

    2024 doi

  10. [18]

    Short- and long-term predictions of chaotic flows and extreme events: a physics-constrained reservoir computing approach

    N. A. K. Doan, W. Polifke, and L. Magri. “Short- and long-term predictions of chaotic flows and extreme events: a physics-constrained reservoir computing approach”. In:Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences477.2253 (Sept. 2021). Pub...

    2021

  11. [19]

    Woojin Cho et al.Parameterized Physics-informed Neural Networks for Parameterized PDEs. Aug. 18, 2024.doi:10.48550/arXiv.2408.09446. arXiv:2408.09446[cs].url:http://arxiv.org/abs/ 2408.09446(visited on 05/27/2025)

    2024 doi arXiv

  12. [20]

    SVD-PINNs: Transfer Learning of Physics- Informed Neural Networks via Singular Value Decomposition

    Yihang Gao, Ka Chun Cheung, and Michael K. Ng. “SVD-PINNs: Transfer Learning of Physics- Informed Neural Networks via Singular Value Decomposition”. In:2022 IEEE Symposium Series on Computational Intelligence (SSCI). Dec. 4, 2022, pp. 1443–1450.doi:10 . 1109 / SSCI51031 . 2022...

    2022

  13. [21]

    Yizheng Wang et al.Transfer Learning in Physics-Informed Neural Networks: Full Fine-Tuning, Lightweight Fine-Tuning, and Low-Rank Adaptation. Feb. 2, 2025.doi:10.48550/arXiv.2502.00782. arXiv:2502.00782[cs].url:http://arxiv.org/abs/2502.00782(visited on 05/28/2025)

    2025 doi arXiv

  14. [22]

    Physics-informed neural network with transfer learning (TL-PINN) based on domain similarity measure for prediction of nuclear reactor transients

    Konstantinos Prantikos et al. “Physics-informed neural network with transfer learning (TL-PINN) based on domain similarity measure for prediction of nuclear reactor transients”. In:Scientific Reports 13.1 (Oct. 6, 2023), p. 16840.issn: 2045-2322.doi:10.1038/s41598- 023- 43325-...

    2023 doi

  15. [23]

    Transfer learning for deep neural network-based partial differential equations solv- ing

    Xinhai Chen et al. “Transfer learning for deep neural network-based partial differential equations solv- ing”. In:Advances in Aerodynamics3.1 (Dec. 8, 2021), p. 36.issn: 2524-6992.doi:10.1186/s42774- 021-00094-7.url:https://doi.org/10.1186/s42774-021-00094-7(visited on 05/28/2025)

    2021 doi

  16. [24]

    Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next

    Salvatore Cuomo et al. “Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next”. In:Journal of Scientific Computing92.3 (Sept. 2022), p. 88.issn: 0885-7474, 1573-7691.doi:10.1007/s10915-022-01939-z.url:https://link.springer.com/10. 1...

    2022 doi

  17. [25]

    arXiv:2103.09655[math].url: http://arxiv.org/abs/2103.09655(visited on 05/28/2025)

    Stefano Markidis.The Old and the New: Can Physics-Informed Deep-Learning Replace Traditional Linear Solvers?July 5, 2021.doi:10.48550/arXiv.2103.09655. arXiv:2103.09655[math].url: http://arxiv.org/abs/2103.09655(visited on 05/28/2025)

    2021 doi arXiv

  18. [26]

    A transfer learning-physics informed neural network (TL-PINN) for vortex- induced vibration

    Hesheng Tang et al. “A transfer learning-physics informed neural network (TL-PINN) for vortex- induced vibration”. In:Ocean Engineering266 (Dec. 2022), p. 113101.issn: 00298018.doi:10 . 1016 / j . oceaneng . 2022 . 113101.url:https : / / linkinghub . elsevier . com / retrieve ...

    2022

  19. [27]

    Rohrhofer et al.On the Role of Fixed Points of Dynamical Systems in Training Physics- Informed Neural Networks

    Franz M. Rohrhofer et al.On the Role of Fixed Points of Dynamical Systems in Training Physics- Informed Neural Networks. Feb. 13, 2023.doi:10.48550/arXiv.2203.13648. arXiv:2203.13648[cs]. url:http://arxiv.org/abs/2203.13648(visited on 06/04/2025)

    2023 doi arXiv

  20. [28]

    Raphael Leiteritz and Dirk Pfl¨ uger.How to Avoid Trivial Solutions in Physics-Informed Neural Net- works. Dec. 10, 2021.doi:10 . 48550 / arXiv . 2112 . 05620. arXiv:2112 . 05620[cs].url:http : //arxiv.org/abs/2112.05620(visited on 06/05/2025). 17

    2021 doi arXiv

  21. [29]

    dPotFit: A computer program to fit diatomic molecule spectral data to potential energy functions

    Robert J. Le Roy. “dPotFit: A computer program to fit diatomic molecule spectral data to potential energy functions”. In:Journal of Quantitative Spectroscopy and Radiative Transfer186 (Jan. 2017), pp. 179–196.issn: 00224073.doi:10.1016/j.jqsrt.2016.06.002.url:https://linkinghu...

    2017 doi

  22. [30]

    LEVEL: A computer program for solving the radial Schr¨ odinger equation for bound and quasibound levels

    Robert J. Le Roy. “LEVEL: A computer program for solving the radial Schr¨ odinger equation for bound and quasibound levels”. In:Journal of Quantitative Spectroscopy and Radiative Transfer186 (Jan. 2017), pp. 167–178.issn: 00224073.doi:10 . 1016 / j . jqsrt . 2016 . 05 . 028.ur...

    2017

  23. [31]

    Transfer learning based physics-informed neural networks for solving inverse problems in engineering structures under different loading scenarios

    Chen Xu et al. “Transfer learning based physics-informed neural networks for solving inverse problems in engineering structures under different loading scenarios”. In:Computer Methods in Applied Mechan- ics and Engineering405 (Feb. 2023), p. 115852.issn: 00457825.doi:10.1016/j...

    2023

  24. [32]

    Transfer Learning-Based Coupling of Smoothed Finite Element Method and Physics-Informed Neural Network for Solving Elastoplastic Inverse Problems

    Meijun Zhou and Gang Mei. “Transfer Learning-Based Coupling of Smoothed Finite Element Method and Physics-Informed Neural Network for Solving Elastoplastic Inverse Problems”. In:Mathematics 11.11 (May 31, 2023), p. 2529.issn: 2227-7390.doi:10.3390/math11112529.url:https://www....

    2023 doi

  25. [33]

    A transfer learning physics-informed deep learning framework for modeling multiple solute dynamics in unsaturated soils

    Hamza Kamil, Azzeddine Soula ¨ ımani, and Abdelaziz Beljadid. “A transfer learning physics-informed deep learning framework for modeling multiple solute dynamics in unsaturated soils”. In:Computer Methods in Applied Mechanics and Engineering431 (Nov. 2024), p. 117276.issn: 004...

    2024

  26. [34]

    May 15, 2023.doi:10

    Shuning Lin and Yong Chen.Gradient-enhanced physics-informed neural networks based on transfer learning for inverse problems of the variable coefficient differential equations. May 15, 2023.doi:10. 48550/arXiv.2305.08310. arXiv:2305.08310[math].url:http://arxiv.org/abs/2305.08...

    2023 doi arXiv

  27. [35]

    Physics-informed Neural Implicit Flow neural network for parametric PDEs

    Zixue Xiang et al. “Physics-informed Neural Implicit Flow neural network for parametric PDEs”. In:Neural Networks185 (May 2025), p. 107166.issn: 08936080.doi:10 . 1016 / j . neunet . 2025 . 107166.url:https://linkinghub.elsevier.com/retrieve/pii/S0893608025000450(visited on 06...

    2025

  28. [36]

    DeepXDE: A Deep Learning Library for Solving Differential Equations

    Lu Lu et al. “DeepXDE: A Deep Learning Library for Solving Differential Equations”. In:SIAM Review 63.1 (Jan. 2021). Publisher: Society for Industrial and Applied Mathematics, pp. 208–228.issn: 0036- 1445.doi:10.1137/19M1274067.url:https://epubs.siam.org/doi/10.1137/19M1274067...

    2021 doi

  29. [37]

    Lena Podina et al.Enhancing Symbolic Regression and Universal Physics-Informed Neural Networks with Dimensional Analysis. Nov. 24, 2024.doi:10.48550/arXiv.2411.15919. arXiv:2411.15919[cs]. url:http://arxiv.org/abs/2411.15919(visited on 07/08/2025)

    2024 doi arXiv

  30. [38]

    A transfer learning enhanced physics-informed neural network for parameter identification in soft materials

    Jing’ang Zhu, Yiheng Xue, and Zishun Liu. “A transfer learning enhanced physics-informed neural network for parameter identification in soft materials”. In:Applied Mathematics and Mechanics45.10 (Oct. 2024). Publisher: Springer Science and Business Media LLC, pp. 1685–1704.iss...

    2024

  31. [39]

    Lazy multivariate higher-order forward-mode AD

    Barak A. Pearlmutter and Jeffrey Mark Siskind. “Lazy multivariate higher-order forward-mode AD”. In:SIGPLAN Not.42.1 (Jan. 17, 2007), pp. 155–160.issn: 0362-1340.doi:10.1145/1190215.1190242. url:https://dl.acm.org/doi/10.1145/1190215.1190242(visited on 07/28/2025)

    2007

  32. [40]

    Junwoo Cho et al.Separable Physics-Informed Neural Networks. Oct. 31, 2023.doi:10 . 48550 / arXiv . 2306 . 15969. arXiv:2306 . 15969[cs].url:http : / / arxiv. org / abs / 2306 . 15969(visited on 07/29/2025)

    2023

  33. [41]

    Wanzhou Lei, Pavlos Protopapas, and Joy Parikh.One-Shot Transfer Learning for Nonlinear ODEs. Nov. 25, 2023.doi:10.48550/arXiv.2311.14931. arXiv:2311.14931[cs].url:http://arxiv.org/ abs/2311.14931(visited on 08/13/2025). 18

    2023 doi arXiv

Pith tools

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