Pith. sign in

REVIEW 3 major objections 6 minor 56 references

Solved in Unit Domain: JacobiNet for Differentiable Coordinate-Transformed PINNs

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read JacobiNet embeds a learned coordinate mapping and its autodiff Jacobian in the same computation graph as the PINN, claiming 15.6x average accuracy gains on irregular-domain PDE benchmarks.

desk verdict A genuinely useful differentiable mapping front-end for PINNs, overclaimed as a general plug-and-play solution; worth a serious referee. read the letter →

arxiv 2508.02537 v3 pith:A2H4PX73 submitted 2025-08-04 cs.LG

Xi Chen , Jianchuan Yang , Junjie Zhang , Runnan Yang , Xu Liu , Hong Wang , Tinghui Zheng , Ziyu Ren
show 1 more author
Wenqi Hu
This is my paper
classification cs.LG MSC 65N3568T07
keywords physics-informedneuralnetworkscoordinatetransformationJacobianlayerautomaticdifferentiationirregulardomainshardboundaryconstraintsNavier-Stokesequationsgeometricgeneralization
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

JacobiNet claims that the longstanding workflow of coordinate-transformed PINNs—mesh the domain, compute the Jacobian numerically, reformulate the PDE by chain rule—can be replaced by a single differentiable graph. A lightweight supervised MLP learns the mapping from physical coordinates $(x,y)$ to a unit reference domain $(\xi,\eta)\in[-1,1]^2$, and automatic differentiation yields the Jacobian $\mathcal{J}$ along with any higher derivatives needed by the PDE residual. Because the mapping and the solver share one computation graph, no mesh, explicit Jacobian storage, or manual PDE reformulation is required. The paper argues this fixes three failure modes of PINNs on irregular geometries—inconsistent normalization, inaccurate boundary enforcement, and imbalanced loss terms—and reports relative $L_2$ errors dropping from 0.11–0.73 to 0.01–0.09 across Laplace, Poisson, Helmholtz, and Navier–Stokes benchmarks, an average 15.6x improvement.

What carries the argument

The load-bearing object is the 'Jacobian Layer': a shallow, tanh-activated MLP placed in front of the PINN, which learns the continuous bijective map $\Phi:(x,y)\mapsto(\xi,\eta)$ and supplies $\mathcal{J}$, $\det\mathcal{J}$, and second-order derivative tensors through automatic differentiation. Because these derivatives are computed on the shared graph, the PDE residual can be formed in physical coordinates while the network sees unit reference coordinates, and reference-domain trial functions like $D(\xi,\eta)=a-\xi$ turn Dirichlet conditions into hard constraints. This single mechanism simultaneously performs the normalization, boundary enforcement, and loss rebalancing that the paper claims.

What would settle it

Take an annulus or a branched vessel-like domain without an analytic mapping, train JacobiNet on any set of user-supplied pairs, and measure the fraction of sampled points with $\det\mathcal{J}\le 0$ (call it $r_{\det,\mathcal{J}}$) plus the maximum normal boundary deviation $\mathcal{E}_{\max}$; if $r_{\det,\mathcal{J}}<100\%$ or $\mathcal{E}_{\max}$ exceeds roughly the $10^{-3}$ scale reported, while the relative $L_2$ error fails to beat the baseline PINN, the central claim of fold-free, boundary-aligned learned mappings for general irregular geometries is false.

Watch

Extended reading notes

Core claim

The central claim is that a neural coordinate transformation can be treated not as preprocessing but as an integral, differentiable layer of the physics solver. For a physical domain $\Omega\subset\mathbb{R}^2$, JacobiNet trains a tanh MLP $\Phi_\theta:(x,y)\mapsto(\xi,\eta)$ under supervised point pairs, with extra weight on boundary points, so boundaries land on coordinate lines of the reference square. The Jacobian $\mathcal{J}=\partial(\xi,\eta)/\partial(x,y)$ is then computed by autograd inside the shared graph, and PDE residuals are evaluated with respect to the original physical coordinates even though the network inputs are reference coordinates. Trial functions built in the reference domain, e.g. $D(\xi,\eta)=a-\xi$ for a Dirichlet edge, enforce boundary conditions exactly and eliminate the boundary loss term. The paper's position is that this combination removes geometric anisotropy from the normalization problem, makes hard boundary constraints practical on complex shapes, and resolves the loss-balance pathology, yielding the reported accuracy gains.

Load-bearing premise

The method's gains depend on having a supplied set of supervised point pairs $(x,y)\mapsto(\xi,\eta)$ per geometry, accurate enough for the MLP to keep boundaries on coordinate lines and the Jacobian determinant positive; if such pairs are unavailable or expensive for arbitrary irregular or multiconnected domains, the claimed plug-and-play generality does not hold.

Editorial extensions

If this is right

  • On U-, S-, and T-shaped domains, the method removes the boundary-loss term while keeping the residual in physical coordinates, so the reported $L_2$ gains come from geometry handling rather than from extra tuning.
  • Once trained on a family of vessel-like geometries, the same mapping network generalizes to unseen shapes, enabling millisecond-scale mapping inference and reducing error by 3.65x on average relative to baseline PINNs.
  • Because the Jacobian comes from autodiff, the approach avoids the truncation and storage-precision losses that the paper attributes to chain-rule reformulations with precomputed Jacobians.
  • JacobiNet is a plug-and-play front-end: the paper shows it can be combined with gradient reweighting and random Fourier feature embeddings, further cutting errors at high Reynolds numbers.
  • The unified graph raises per-epoch cost by roughly 16% for scalar problems and about 50% for 3D Navier-Stokes, but requires far fewer epochs to reach a target accuracy, yielding net training-time savings.

Reading between the lines

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

  • A natural stress test is a multiconnected domain such as an annulus or a branching vessel, where the paper explicitly leaves supervised-data construction and topological consistency as open challenges; whether the method keeps $r_{\det,\mathcal{J}}=100\%$ there is not yet shown.
  • If the Jacobian layer is reused as a normalization front-end for neural operators, its differentiable geometry encoding could give them the same boundary-aligned input space; the paper mentions this possibility only briefly.
  • The reported improvements bundle the learned mapping and the hard-constraint trial functions; an ablation that isolates the mapping's contribution to loss balance in a domain where distance functions are analytically known would clarify which ingredient carries the gain.
  • The learned mapping's boundary alignment is currently checked by post-hoc metrics; making the boundary-alignment loss a training-time constraint could broaden the method to more complex topologies.
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 / 6 minor

Summary. The manuscript proposes JacobiNet, a framework that prepends a supervised MLP coordinate transformation to a PINN. The mapping network is trained on user-supplied interior and boundary point pairs (x,y)->(xi,eta) via Eq. (18), and is inserted into the PINN computational graph so that physical-space derivatives are obtained by automatic differentiation through the mapping; no stored Jacobian or manual chain-rule PDE reformulation is needed. Dirichlet conditions are imposed with trial functions built from mapped coordinates (Eq. 24). Experiments cover Laplace on a U-shape, Poisson on an S-shape, Helmholtz on a T-shape, 3D Navier-Stokes in stenosed vessels, and parametric 2D vessel families, reporting relative L2 reductions and ablations against coordinate-transformed-only PINNs and several baselines.

Significance. The central mechanism is sound and the ablation design is informative: comparing against coordinate-transformed-only PINNs isolates the effect of hard constraints, and the reported error reductions are large and consistent in most configurations. The per-epoch cost is honestly reported in Table 2, and the parametric vessel experiment addresses a practically relevant use case. If the claims are narrowed to settings where a target mapping is available or hand-constructible, this is a useful contribution to geometry handling in PINNs. There is no circularity: the PDE residual loss is evaluated against the governing equations, and the mapping network is trained on user-provided pairs independent of the PDE solution. However, the paper's headline generality is not supported by the experiments, because the mapping network only fits user-provided pairs and no automatic construction of transformations is proposed. The lack of code or data release also prevents independent verification.

major comments (3)
  1. [Section 3.1, Eq. (18); Section 5] The abstract and Introduction promise a general, mesh-free, plug-and-play replacement for irregular and multiconnected domains, but the method requires the user to supply target coordinate pairs (xi,eta) for every geometry. In all experiments the target pairs come from hand-designed constructions (centerline unfolding in Section 4.1, radial normalization in Section 4.4, slice-wise min-max in Section 4.3) or from a parametric template in Section 4.5; Eq. (18) only regresses those pairs. The concluding section explicitly concedes that "the construction of supervised data and the preservation of topological consistency remain open challenges." Consequently, the claim that JacobiNet avoids case-specific geometric preprocessing is not supported. The authors should either provide an automated way to construct the target transformation or restrict the claims and title-level promises to the demonstrated setting of known or hand-constructible mappings.
  2. [Section 3.3, Eq. (24)] The hard-constraint mechanism is only as exact as the learned boundary mapping. Because Phi is a regressor trained on finite samples, mapped boundary points do not land exactly on the coordinate lines xi=+/-1; the paper's own metrics in Section 4.1 report RMSE_bd=1.8 mm and E_max=5.0 mm, and Appendix-2 reports E_max up to 8e-3. Thus D_u(xi,eta) does not vanish exactly on the physical boundary, so the Dirichlet condition is satisfied only approximately. The text should qualify "hard constraints" accordingly and, ideally, report the resulting physical-space boundary residual or propagate the mapping error to the solution error.
  3. [Section 4.7, Tables 3-4] The text states that JacobiNet "consistently demonstrates superior performance, particularly at higher Reynolds numbers (Re > 100)." This is contradicted by the low-Re rows: in Table 3 at Re=10, JacobiNet has relative L2 errors 0.028/0.032 versus 0.020/0.022 for RFF, and in Table 4 at Re=10 the velocity error is 0.033 versus 0.022 for RFF. The superiority claim should be restricted to Re>=100 (or to the average over the tested range), and the borderline cases should be discussed rather than glossed over.
minor comments (6)
  1. [Section 4.1] The reported value "RMSE_bd=1.8m" appears to be a typo for "1.8 mm"; please correct all unit inconsistencies in the mapping accuracy metrics.
  2. [Section 4.5] The text refers to Fig. 5A and Fig. 5C for the feature extraction module and parametric mapping, but the parametric vessel framework is actually presented in Fig. 6; the cross-references should be fixed.
  3. [Section 2, Eq. (12)] The tensor contraction in Eq. (12) is notationally ambiguous; a component-wise statement of the Hessian transformation would help readers verify the chain-rule implementation.
  4. [Tables 1, 3, and 4] All reported relative L2 errors are point estimates with no variance across seeds; given the stochasticity of PINN training, reporting mean and standard deviation over multiple runs would strengthen the quantitative claims.
  5. [Data availability] The code and datasets are promised only "upon publication"; for a methods paper, a working repository or detailed pseudocode should be provided for independent verification.
  6. [Section 5] The statement that JacobiNet introduces "over 20% additional cost" is inconsistent with Table 2, where the 2D cases show a 16-18% increase on average; the phrasing should be aligned with the data.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: JacobiNet's mapping network is supervised by user-supplied point pairs and the PINN loss is the PDE residual, so the accuracy gains are not forced by construction.

full rationale

The claimed derivation chain is self-contained and non-circular. The mapping network is trained by supervised regression on user-provided point pairs (x,y)->(xi,eta) via Eq. (18); these targets are supplied by hand-designed or template-based transformations, not by the PDE solution. The PINN loss is the physical PDE residual evaluated through autodiff Jacobians (Eq. (16)), and the reported relative L2 errors are measured against independent FVM reference solutions (Eq. (28)), not against a fitted constant or the training objective. Boundary conditions are enforced by trial functions constructed in the reference domain (Eq. (24)), which is a mechanism that relies on an accurate learned mapping but does not itself define the mapping in terms of the PDE output. The paper's acknowledged limitation in Section 5 - that 'construction of supervised data and the preservation of topological consistency remain open challenges' - is an external dependency on available target transformations, not an equivalence between inputs and predictions. Self-citations to refs. [43] and [47] appear only for physiological parameter choices and a comparison baseline, and are not load-bearing for the central claim. No fitted parameter is relabeled as a prediction, no uniqueness theorem is imported, and no ansatz is smuggled in via citation. The central accuracy comparison therefore has independent content, and the score is 0.

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

The method's central claim rests on user-supplied mapping supervision, manual construction of trial functions, and standard assumptions about neural network approximation and FVM ground truth. No new physics is postulated.

free parameters (6)
  • Boundary supervision weight lambda = 10
    Chosen from the sensitivity sweep in Appendix-2; lambda in [5,20] gives the lowest PDE L2 error. It controls boundary mapping accuracy and final solution accuracy.
  • JacobiNet architecture = 2 hidden layers x 128 neurons, tanh
    Chosen empirically and not justified by analysis. The mapping approximation quality and the autodiff Jacobians depend on it.
  • PINN architecture and training schedule = 3-4 layers x 64 neurons, Adam lr 1e-3 to 1e-5, 1k-80k epochs
    Per-case settings in Appendix-3, chosen empirically without search. The reported accuracies are conditional on these choices.
  • Trial function forms for hard boundary conditions = e.g., phi_wall=1-(xi^2+eta^2), phi_in=(1-zeta)/2, D_u=(a-xi) or (1-exp(a-xi))
    Manually designed per geometry. They determine whether boundary conditions are exactly met and hence the size of the accuracy gain.
  • Ground-truth mapping operations = unfolding along centerline, radial normalization, y-slice min-max normalization
    The user must define these operations; JacobiNet regresses onto them. The quality and bijectivity of the learned mapping inherit from these choices.
  • Feature extraction module for parametric geometries = Not specified in detail
    Encodes vessel length and deformation to disambiguate geometries in Section 4.5; its hyperparameters are not given, but the generalization results depend on it.
assumptions (6)
  • domain assumption For each domain there exists a smooth, orientation-preserving bijection to the unit reference domain.
    Section 3.1 assumes det J > 0 and verifies it post-hoc; without such a mapping the learned network cannot be bijective.
  • domain assumption Supervised point pairs (x_i, y_i) to (xi_i, eta_i) faithfully represent the desired mapping.
    Eq. (18) trains on these pairs; the PDE solution accuracy inherits the quality of this supervision.
  • standard math Neural networks of the chosen size can approximate the target mapping and PDE solution sufficiently well.
    Universal approximation is invoked implicitly; the experiment sizes (2x128 and 3-4x64) are not proven sufficient.
  • domain assumption Reference-domain trial functions constructed from the mapped coordinates exactly enforce the prescribed Dirichlet conditions.
    Eqs. (24), (40)-(45) rely on boundaries aligning with coordinate lines in the mapped domain; this holds for the hand-designed mappings but is not automatic.
  • domain assumption FVM/CFD solutions on high-fidelity meshes are accurate references for the PDEs.
    All L2 errors are measured against ANSYS Fluent FVM solutions (Sections 4.4-4.5).
  • domain assumption The PDE residual optimization with Adam reaches a good minimum for the composed network.
    No convergence guarantee is given; accuracy claims depend on successful training of the combined graph.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Solved in Unit Domain: JacobiNet for Differentiable Coordinate-Transformed PINNs." pith.science (2026). https://pith.science/paper/A2H4PX73

@misc{pith2026250802537,
  author       = {Pith},
  title        = {Pith review of: Solved in Unit Domain: JacobiNet for Differentiable Coordinate-Transformed PINNs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2H4PX73}},
  note         = {Machine review of arXiv:2508.02537}
}
read the original abstract

Physics-Informed Neural Networks (PINNs) offer a powerful framework for solving PDEs by embedding physical laws into the learning process. However, when applied to domains with irregular boundaries, PINNs often suffer from instability and slow convergence, which stems from (1) inconsistent normalization due to geometric anisotropy, (2) inaccurate boundary enforcement, and (3) imbalanced loss term competition. A common workaround is to map the domain to a regular space. Yet, conventional mapping methods rely on case-specific meshes, define Jacobians at pre-specified fixed nodes, reformulate PDEs via the chain rule-making them incompatible with modern automatic differentiation, tensor-based frameworks. To bridge this gap, we propose JacobiNet, a learning-based coordinate-transformed PINN framework that unifies domain mapping and PDE solving within an end-to-end differentiable architecture. JacobiNet enables direct Jacobian computation via autograd, shares computation graph with downstream PINNs, thereby avoiding case-specific meshing, explicit Jacobian computation/storage, and manual PDE reformulation while unlocking geometric-editing operations. Separating physical modeling from geometric complexity, JacobiNet (1) addresses normalization challenges in the original anisotropic coordinates, (2) facilitates the hard enforcement of boundary conditions, and (3) mitigates the long-standing imbalance among loss terms. Evaluated on various PDEs, JacobiNet reduces the relative L2 error from 0.11-0.73 to 0.01-0.09, achieving an average 15.6x improvement in accuracy. In vessel-like domains with varying shapes, JacobiNet enables millisecond-level mapping inference for unseen geometries, improves prediction accuracy by an average of 3.65x, while delivering over 10x speedup-demonstrating strong generalization, accuracy, and efficiency.

Figures

Figures reproduced from arXiv: 2508.02537 by the authors.

Figure 1
Figure 1. Standard & coordinate-transformed PINN workflows in complex domains and challenges. (A)Standard [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Learning-based JacobiNet enable continuous, differentiable mappings and geometric editing operations, [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. JacobiNet + PINN: End-to-end differentiable integration of domain mapping and PDE solving. [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: JacobiNet learns diverse mapping across complex domains and enables accurate PDE solutions. (A) [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: JacobiNet performance in 3D stenosed vessels. (A) Concentric and eccentric stenosed geometries with [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]
Figure 6
Figure 6. Figure 6: JacobiNet provides a generalizable framework that maps vessel-like geometries with varying lengths [PITH_FULL_IMAGE:figures/full_fig_p019_6.png]
Figure 7
Figure 7. Figure 7: Quantitative comparison of JacobiNet and baseline & state-of-the-art methods. (A) Improvement in [PITH_FULL_IMAGE:figures/full_fig_p021_7.png]
Figure 8
Figure 8. Figure 8: Experimental validation of loss-term imbalance in complex geometries. (A) Domains of increasing [PITH_FULL_IMAGE:figures/full_fig_p029_8.png]
Figure 9
Figure 9. Figure 9: Sensitivity analysis of boundary weight 𝜆. (A) Root mean square error of interior points, 𝑅𝑀𝑆𝐸௜௡; (B) Root mean square error of boundary points, 𝑅𝑀𝑆𝐸௕ௗ; (C) Maximum normal deviation of boundary mapped points ℰ௠௔௫; (D) Final PDE prediction relative 𝐿ଶ error [PITH_FULL_…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 52 canonical work pages

  1. [1]

    Perdikaris, and G.E

    Raissi, M., 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. Journal of Computational physics, 2019. 378: p. 686-707

    work page 2019

  2. [2]

    Journal of Machine Learning Research, 2018

    Raissi, M., Deep hidden physics models: Deep learning of nonlinear partial differential equations. Journal of Machine Learning Research, 2018. 19(25): p. 1-24

    work page 2018

  3. [3]

    Nature Reviews Physics, 2021

    Karniadakis, G.E., et al., Physics-informed machine learning. Nature Reviews Physics, 2021. 3(6): p. 422- 440

    work page 2021

  4. [4]

    Yazdani, and G.E

    Raissi, M., A. Yazdani, and G.E. Karniadakis, Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations. Science, 2020. 367(6481): p. 1026-1030

    work page 2020

  5. [5]

    Computer Methods in Applied Mechanics and Engineering, 2023

    Diao, Y ., et al., Solving multi-material problems in solid mechanics using physics-informed neural networks based on domain decomposition technology. Computer Methods in Applied Mechanics and Engineering, 2023. 413: p. 116120

    work page 2023

  6. [6]

    Inverse Problems, 2023

    Tanyu, D.N., et al., Deep learning methods for partial differential equations and related parameter identification problems. Inverse Problems, 2023. 39(10): p. 103001

    work page 2023

  7. [7]

    SIAM Journal on Scientific Computing, 2021

    Lu, L., et al., Physics-informed neural networks with hard constraints for inverse design. SIAM Journal on Scientific Computing, 2021. 43(6): p. B1105-B1132

    work page 2021

  8. [8]

    Sengupta, and M

    Sun, Y ., U. Sengupta, and M. Juniper, Physics-informed deep learning for simultaneous surrogate modeling and PDE-constrained optimization of an airfoil geometry. Computer Methods in Applied Mechanics and Engineering, 2023. 411: p. 116042

    work page 2023

Show all 56 references
  1. [9]

    Journal of Scientific Computing, 2022

    Cuomo, S., et al., Scientific machine learning through physics–informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 2022. 92(3): p. 88

    2022

  2. [10]

    Markidis, S., The old and the new: Can physics-informed deep-learning replace traditional linear solvers? Frontiers in big Data, 2021. 4: p. 669097

    2021

  3. [11]

    Advances in neural information processing systems, 2020

    Tancik, M., et al., Fourier features let networks learn high frequency functions in low dimensional domains. Advances in neural information processing systems, 2020. 33: p. 7537-7547

    2020

  4. [12]

    Wang, and P

    Wang, S., H. Wang, and P. Perdikaris, On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 2021. 384: p. 113938

    2021

  5. [13]

    Jagtap, A.D. 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. Communications in Computational Physics, 2020. 28(5)

    2020

  6. [14]

    Kashefi, A. and T. Mukerji, Physics-informed PointNet: A deep learning solver for steady-state incompressible flows and thermal fields on multiple sets of irregular geometries. Journal of Computational Physics, 2022. 468: p. 111510

    2022

  7. [15]

    Sun, and J.-X

    Gao, H., L. Sun, and J.-X. Wang, PhyGeoNet: Physics-informed geometry-adaptive convolutional neural networks for solving parameterized steady-state PDEs on irregular domain. Journal of Computational Physics,

  8. [16]

    Nature Computational Science, 2024

    Yin, M., et al., A scalable framework for learning the geometry-dependent solution operators of partial differential equations. Nature Computational Science, 2024. 4(12): p. 928-940

    2024

  9. [17]

    Journal of Machine Learning Research, 2023

    Li, Z., et al., Fourier neural operator with learned deformations for pdes on general geometries. Journal of Machine Learning Research, 2023. 24(388): p. 1-26

    2023

  10. [18]

    Gal-Chen, T. and R.C. Somerville, On the use of a coordinate transformation for the solution of the Navier- Stokes equations. Journal of Computational Physics, 1975. 17(2): p. 209-228

    1975

  11. [19]

    Suh, and J.B

    Kang, D.-G., D.C. Suh, and J.B. Ra, Three-dimensional blood vessel quantification via centerline deformation. IEEE Transactions on Medical Imaging, 2008. 28(3): p. 405-414

    2008

  12. [20]

    Blood vessel segmentation and centerline tracking using local structure analysis

    Kumar, R.P., et al. Blood vessel segmentation and centerline tracking using local structure analysis. in 6th European Conference of the International Federation for Medical and Biological Engineering: MBEC 2014, 7- 11 September 2014, Dubrovnik, Croatia. 2015. Springer

    2014

  13. [21]

    Sukumar, N. and A. Srivastava, Exact imposition of boundary conditions with distance functions in physics- informed deep neural networks. Computer Methods in Applied Mechanics and Engineering, 2022. 389: p. 114333

    2022

  14. [22]

    Teng, and P

    Wang, S., Y . Teng, and P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics- informed neural networks. SIAM Journal on Scientific Computing, 2021. 43(5): p. A3055-A3081

    2021

  15. [23]

    Yu, and P

    Wang, S., X. Yu, and P. Perdikaris, When and why PINNs fail to train: A neural tangent kernel perspective. Journal of Computational Physics, 2022. 449: p. 110768

    2022

  16. [24]

    Computer Methods in Applied Mechanics and Engineering, 2023

    Wu, C., et al., A comprehensive study of non-adaptive and residual-based adaptive sampling for physics- informed neural networks. Computer Methods in Applied Mechanics and Engineering, 2023. 403: p. 115671

    2023

  17. [25]

    arXiv preprint arXiv:2207.02338, 2022

    Daw, A., et al., Rethinking the importance of sampling in physics-informed neural networks. arXiv preprint arXiv:2207.02338, 2022

    2022 arXiv

  18. [26]

    McClenny, L.D. and U.M. Braga-Neto, Self-adaptive physics-informed neural networks. Journal of Computational Physics, 2023. 474: p. 111722

    2023

  19. [27]

    Communications in Mathematics and Statistics, 2023

    Zhan, Z., et al., Boundary correspondence for iso-geometric analysis based on deep learning. Communications in Mathematics and Statistics, 2023. 11(1): p. 131-150

    2023

  20. [28]

    Chen, and W

    Pan, M., F. Chen, and W. Tong, V olumetric spline parameterization for isogeometric analysis. Computer Methods in Applied Mechanics and Engineering, 2020. 359: p. 112769

    2020

  21. [29]

    Wang, and F

    Zhan, Z., W. Wang, and F. Chen, Simultaneous boundary and interior parameterization of planar domains via deep learning. Computer-Aided Design, 2024. 166: p. 103621

    2024

  22. [30]

    Wang, and F

    Zhan, Z., W. Wang, and F. Chen, Integral parameterization of volumetric domains via deep neural networks. Computer Methods in Applied Mechanics and Engineering, 2025. 441: p. 117988

    2025

  23. [31]

    Neural Computing and Applications, 2024

    Giannelli, C., et al., BIDGCN: boundary-informed dynamic graph convolutional network for adaptive spline fitting of scattered data. Neural Computing and Applications, 2024. 36(28): p. 17261-17284

    2024

  24. [32]

    Wang, and F

    Zhan, Z., W. Wang, and F. Chen, Fast parameterization of planar domains for isogeometric analysis via generalization of deep neural network. Computer Aided Geometric Design, 2024. 111: p. 102313

    2024

  25. [33]

    Journal of Computational Physics, 2025: p

    Xu, S., et al., On the preprocessing of physics-informed neural networks: how to better utilize data in fluid mechanics. Journal of Computational Physics, 2025: p. 113837

    2025

  26. [34]

    Glorot, X. and Y . Bengio. Understanding the difficulty of training deep feedforward neural networks. in Proceedings of the thirteenth international conference on artificial intelligence and statistics. 2010. JMLR Workshop and Conference Proceedings

    2010

  27. [35]

    Basir, S. and I. Senocak. Critical investigation of failure modes in physics-informed neural networks. in AiAA SCITECH 2022 Forum. 2022

    2022

  28. [36]

    Advances in neural information processing systems, 2021

    Krishnapriyan, A., et al., Characterizing possible failure modes in physics-informed neural networks. Advances in neural information processing systems, 2021. 34: p. 26548-26560

    2021

  29. [37]

    Advances in Neural Information Processing Systems, 2022

    Liu, S., et al., A unified hard-constraint framework for solving geometrically complex pdes. Advances in Neural Information Processing Systems, 2022. 35: p. 20287-20299

    2022

  30. [38]

    Gradient descent only converges to minimizers

    Lee, J.D., et al. Gradient descent only converges to minimizers. in Conference on learning theory. 2016. PMLR

    2016

  31. [39]

    Neurocomputing, 2022

    Xiang, Z., et al., Self-adaptive loss balanced physics-informed neural networks. Neurocomputing, 2022. 496: p. 11-34

    2022

  32. [40]

    Computer Methods in Applied Mechanics and Engineering, 2024

    Yang, J., et al., Adaptive task decomposition physics-informed neural networks. Computer Methods in Applied Mechanics and Engineering, 2024. 418: p. 116561

    2024

  33. [41]

    Journal of Biomechanics, 2025

    Cao, H., et al., Standardizing plaque impact on overall coronary hemodynamics using a binomial equation based on the equivalent energy loss. Journal of Biomechanics, 2025. 179: p. 112461

    2025

  34. [42]

    Heart, Lung and Circulation,

    Sakakura, K., et al., Pathophysiology of atherosclerosis plaque progression. Heart, Lung and Circulation,

  35. [43]

    Frontiers in Bioengineering and Biotechnology, 2024

    Chen, X., et al., Hemodynamic influence of mild stenosis morphology in different coronary arteries: a computational fluid dynamic modelling study. Frontiers in Bioengineering and Biotechnology, 2024. 12: p. 1439846

    2024

  36. [44]

    Surgical and Radiologic Anatomy, 2023

    Javed, S., et al., Multi-slice CT analysis of the length of left main coronary artery: its relation to sex, age, diameter and branching pattern of left main coronary artery, and coronary dominance. Surgical and Radiologic Anatomy, 2023. 45(8): p. 1009-1019

    2023

  37. [45]

    JACC: cardiovascular imaging, 2016

    Cury, R.C., et al., Coronary artery disease-reporting and data system (CAD-RADS) an expert consensus document of SCCT, ACR and NASCI: endorsed by the ACC. JACC: cardiovascular imaging, 2016. 9(9): p. 1099-1113

    2016

  38. [46]

    arXiv preprint arXiv:2308.08468, 2023

    Wang, S., et al., An expert's guide to training physics-informed neural networks. arXiv preprint arXiv:2308.08468, 2023

    2023 arXiv

  39. [47]

    Computer Methods in Applied Mechanics and Engineering, 2025

    Chen, X., et al., Hemodynamics modeling with physics-informed neural networks: A progressive boundary complexity approach. Computer Methods in Applied Mechanics and Engineering, 2025. 438: p. 117851

    2025

  40. [48]

    arXiv preprint arXiv:2010.08895, 2020

    Li, Z., et al., Fourier neural operator for parametric partial differential equations. arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  41. [49]

    Nature machine intelligence, 2021

    Lu, L., et al., Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nature machine intelligence, 2021. 3(3): p. 218-229

    2021

  42. [50]

    Jin, and G.E

    Lu, L., P . Jin, and G.E. Karniadakis, Deeponet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators. arXiv preprint arXiv:1910.03193, 2019

    1910 arXiv

  43. [51]

    ACM/JMS Journal of Data Science, 2024

    Li, Z., et al., Physics-informed neural operator for learning partial differential equations. ACM/JMS Journal of Data Science, 2024. 1(3): p. 1-27

    2024

  44. [52]

    IEEE transactions on neural networks and learning systems, 2020

    Wu, Z., et al., A comprehensive survey on graph neural networks. IEEE transactions on neural networks and learning systems, 2020. 32(1): p. 4-24

    2020

  45. [53]

    arXiv preprint arXiv:2411.08122, 2024

    Zhou, C., et al., Physics-Informed Neural Networks with Complementary Soft and Hard Constraints for Solving Complex Boundary Navier-Stokes Equations. arXiv preprint arXiv:2411.08122, 2024. Appendix-1 Experimental validation of loss-term imbalance in complex geometries To quant...

    2024 arXiv

  46. [54]

    The definitions of these loss terms are given in Eqs

    Loss-ratio curves: ℛ௟௢௦௦(𝑡) =ℒೠ(௧) ℒ್(௧), (3) where ℛ௟௢௦௦(𝑡) is the ratio between the PDE residual loss ℒ௨ and the boundary condition loss ℒ௕. The definitions of these loss terms are given in Eqs. (5)-(6) of the main text. Here, 𝑡 denotes the training iteration. The loss-ratio...

  47. [55]

    This metric reflects the relative influence of each loss term on parameter updates

    Gradient-norm ratio: ℛ௚௥௔ௗ(𝑡)=‖∇ഛℒೠ(௧)‖మ ‖∇ഛℒ್(௧)‖మ , (4) where ℛ௚௥௔ௗ(𝑡) computes the ratio of the gradient norms of the loss terms with respect to the trainable parameters 𝜗. This metric reflects the relative influence of each loss term on parameter updates

  48. [56]

    The performance of PINNs across domains is shown in Fig

    Relative 𝐿ଶ error 𝐿ଶ,୰ୣ୪= ට∑ [௨(௫೔,௬೔)ି௨ෝ(௫೔,௬೔)]మಿ೔సభ ට∑ [௨(௫೔,௬೔)]మಿ೔సభ , (5) The final accuracy is measured by the relative 𝐿ଶ error between the predicted solution 𝑢ො and the ground truth solution 𝑢. The performance of PINNs across domains is shown in Fig. 8C. In the unit d...

Pith tools

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