Spatio-Temporal Uncertainty-Modulated Physics-Informed Neural Networks for Solving Hyperbolic Conservation Laws with Strong Shocks

Darui Zhao; Fujun Liu; Ze Tao

arxiv: 2604.00029 · v2 · pith:JHXRYHZNnew · submitted 2026-03-20 · ⚛️ physics.comp-ph

Spatio-Temporal Uncertainty-Modulated Physics-Informed Neural Networks for Solving Hyperbolic Conservation Laws with Strong Shocks

Darui Zhao , Ze Tao , Fujun Liu This is my paper

Pith reviewed 2026-05-25 06:21 UTC · model grok-4.3

classification ⚛️ physics.comp-ph

keywords Physics-informed neural networksUncertainty modulationHyperbolic conservation lawsShock wavesComputational fluid dynamicsMesh-free methodsMulti-task learning

0 comments

The pith

UM-PINN reinterprets PINN training as uncertainty-modulated multi-task learning to resolve strong shocks in hyperbolic conservation laws.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes UM-PINN to overcome the failure of standard Physics-Informed Neural Networks at capturing shock waves in compressible flows, a failure tied to gradient pathology at discontinuities. It reinterprets training as multi-task learning under homoscedastic aleatoric uncertainty, using a gradient-based spatial mask and learnable variance parameters to dynamically weight PDE residuals against initial conditions across space and time. Quasi-Monte Carlo Sobol sampling further stabilizes the process. Validation on the Sod shock tube, Shu-Osher problem, and 2D Riemann interaction shows orders-of-magnitude gains in accuracy and shock sharpness over baselines. A sympathetic reader would care because the approach offers a mesh-free route to simulating hyperbolic systems where conventional weighting schemes break down.

Core claim

By reinterpreting the training process as a multi-task learning problem governed by homoscedastic aleatoric uncertainty and integrating a gradient-based spatial mask with learnable variance parameters, the UM-PINN framework dynamically balances the contributions of PDE residuals and initial conditions across the spatiotemporal domain, achieving robust resolution of strong shocks in hyperbolic conservation laws that standard gradient-based weighting schemes cannot handle.

What carries the argument

The Spatio-Temporal Uncertainty-Modulated PINN (UM-PINN) that uses homoscedastic aleatoric uncertainty with learnable variance parameters and a gradient-based spatial mask to balance loss terms.

If this is right

UM-PINN achieves orders of magnitude improvement in accuracy and shock resolution on the 1D Sod shock tube.
UM-PINN captures high-frequency oscillations in the Shu-Osher problem where standard methods fail.
UM-PINN accurately handles complex wave interactions in the 2D Riemann problem.
UM-PINN establishes a robust mesh-free paradigm for Computational Fluid Dynamics on hyperbolic systems.

Where Pith is reading between the lines

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

The uncertainty-modulation approach could extend to other discontinuous hyperbolic problems such as contact discontinuities.
Learnable variance parameters might be adapted for adaptive weighting in long-time integration of unsteady flows.
The framework could reduce reliance on body-fitted meshes when simulating shocks in complex geometries.

Load-bearing premise

That reinterpreting training as multi-task learning with homoscedastic aleatoric uncertainty and a gradient-based mask will balance PDE residuals and initial conditions without introducing new instabilities or biases at discontinuities.

What would settle it

If UM-PINN applied to the 1D Sod shock tube fails to produce orders-of-magnitude lower error or sharper shock profiles than baseline PINNs, the central performance claim would be falsified.

Figures

Figures reproduced from arXiv: 2604.00029 by Darui Zhao, Fujun Liu, Ze Tao.

**Figure 1.** Figure 1: Schematic architecture of the Uncertainty-Modulated Physics-Informed Neural Network (UM-PINN). The network predicts the primitive variables (𝜌, 𝑢, 𝑝) from spatio-temporal coordinates (𝑡, 𝑥). The core innovation is the Uncertainty-Modulated (UM) Module simultaneously optimize multiple competing physical constraints, we seek to maximize the joint log-likelihood. For independent tasks, minimizing the negativ… view at source ↗

**Figure 2.** Figure 2: Comparison of predicted profiles for density, velocity, and pressure against analytical solutions for the 1D Sod shock tube at 𝑡 = 0.5 [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Quantitative error metrics (RMSE, Relative 𝐿2 error, and 𝐿∞ error) comparing the Baseline PINN and the proposed UM-PINN for the Sod shock tube problem. 3.2. 1D Shu-Osher Problem We proceed to the 1D Shu-Osher problem, a rigorous benchmark designed to evaluate the solver’s ability to capture high-frequency flow features. The problem describes the interaction between a moving shock wave (𝑀 = 3) and a sinuso… view at source ↗

**Figure 4.** Figure 4: Comparative analysis of the Shu-Osher problem at 𝑡 = 1.80. (a) The proposed UM-PINN (left) successfully resolves the intricate oscillations and matches the reference solution with high accuracy. (b)The Standard Baseline PINN (right) fails to capture high-frequency entropy waves due to spectral bias. both the correct amplitude and phase which are critical for high-speed aero-thermodynamics simulations. 3.3.… view at source ↗

**Figure 5.** Figure 5: Qualitative comparison of the steady-state Riemann problem results. (a) Density (𝜌) fields; (b) Pressure (𝑝) fields. In each panel, the top row displays the sharp interfaces captured by the proposed UM-PINN, while the bottom row shows the results from the Baseline method, which suffer from numerical diffusion [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗

**Figure 6.** Figure 6: Error metric comparison for the 2D Riemann problem across all conserved (𝜌, 𝜌𝑢, 𝜌𝑣, 𝜌𝐸) and primitive (𝜌, 𝑢, 𝑣, 𝑝) variables. the topological consistency of complex 2D shock interactions. Unlike the Baseline which exhibits rounded shock corners and blurred slip lines, the UM-PINN maintains sharp interfaces. This is because the spatial modulation acts as a localized regularizer, focusing the network’s capa… view at source ↗

**Figure 7.** Figure 7: Training instability and failure modes of GradNorm and LRA algorithms when applied to the Shu-Osher problem. The methods often converge to trivial solutions or fail to resolve the high-frequency entropy waves behind the shock [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗

**Figure 8.** Figure 8: Comparative performance of LRA and GradNorm on the 1D Sod shock tube, highlighting non-physical artifacts (e.g., Gibbs-like oscillations) and convergence issues in capturing the shock front and contact discontinuity. In contrast to the fragility of LRA and GradNorm, UMPINN achieves stable and accurate convergence on the exact same problems as demonstrated in Section 3. The key advantage lies in the mecha… view at source ↗

read the original abstract

Physics-Informed Neural Networks (PINNs) frequently encounter difficulties in accurately resolving shock waves within high-speed compressible flows, a failure largely attributed to the "gradient pathology" arising from extreme stiffness at discontinuities. To overcome this limitation, we propose the Spatio-Temporal Uncertainty-Modulated PINN (UM-PINN), a probabilistic framework that reinterprets the training process as a multi-task learning problem governed by homoscedastic aleatoric uncertainty. By integrating a gradient-based spatial mask with learnable variance parameters, our method dynamically balances the conflicting contributions of Partial Differential Equation (PDE) residuals and initial conditions across the spatiotemporal domain, further stabilized by Quasi-Monte Carlo Sobol sampling. We validate the framework against challenging benchmarks, including the one-dimensional (1D) Sod shock tube, the high-frequency Shu-Osher problem, and the complex two-dimensional (2D) Riemann interaction, where standard gradient-based weighting schemes typically fail. Experimental results demonstrate that UM-PINN achieves orders of magnitude improvement in accuracy and shock resolution compared to baseline methods, establishing a robust new paradigm for mesh-free Computational Fluid Dynamics in hyperbolic systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

UM-PINN tries to fix PINN shock failures with homoscedastic uncertainty weighting plus a gradient mask, but the abstract's big accuracy claims rest on an assumption that may not hold at discontinuities.

read the letter

The main point is that this paper presents UM-PINN as a way to handle strong shocks in hyperbolic conservation laws by recasting the loss as multi-task learning under homoscedastic aleatoric uncertainty, with a gradient-based spatial mask and learnable variances to balance PDE residuals against initial conditions, plus Sobol sampling. It tests on Sod, Shu-Osher, and 2D Riemann cases and says it gets orders-of-magnitude better accuracy than baselines where standard weighting fails. That framing is new enough in the PINN-for-CFD space and directly targets a documented failure mode. The approach is reasonable on its face and gives credit to the underlying gradient pathology problem. What it does well is offer a probabilistic angle that could let the network down-weight stiff regions without manual tuning. The soft spot is the homoscedastic assumption itself. Shock problems have residuals that jump by orders of magnitude across the discontinuity, so a single variance per task is unlikely to capture the spatial heterogeneity; the balancing may collapse to something closer to static weighting. The abstract states the gains but supplies no equations for the modulated loss, no error tables, no ablation on the mask versus the uncertainty terms, and no discussion of whether the variances actually vary meaningfully near shocks. Without those, the central claim stays hard to evaluate. This is for people already working on PINNs for compressible flows who want to try uncertainty-based reweighting. A reader who needs reproducible improvements for engineering CFD would get limited value until the experiments are shown in detail. It deserves a serious referee because the target problem is real and the method is a coherent extension, even if the evidence needs strengthening.

Referee Report

2 major / 1 minor

Summary. The paper proposes Spatio-Temporal Uncertainty-Modulated Physics-Informed Neural Networks (UM-PINN) for solving hyperbolic conservation laws with strong shocks. It reinterprets PINN training as multi-task learning under homoscedastic aleatoric uncertainty, augmented by a gradient-based spatial mask, learnable variance parameters, and Quasi-Monte Carlo Sobol sampling to dynamically balance PDE residuals and initial conditions. The framework is tested on the 1D Sod shock tube, high-frequency Shu-Osher problem, and 2D Riemann interaction, with the central claim that it achieves orders-of-magnitude gains in accuracy and shock resolution over standard gradient-based weighting schemes.

Significance. If the experimental gains are reproducible and the balancing mechanism is robust, the approach could provide a practical mesh-free alternative for high-speed compressible flows where conventional PINNs fail at discontinuities. The explicit use of uncertainty modulation and Sobol sampling for stabilization is a concrete technical contribution that could be adopted more broadly if the homoscedastic assumption holds.

major comments (2)

Abstract and method description: the central claim of dynamic balancing rests on a single learnable variance parameter per task (homoscedastic aleatoric uncertainty) plus a gradient-based spatial mask. In hyperbolic systems the residual magnitude and its statistics jump by orders of magnitude across shocks; a spatially constant variance cannot capture this heterogeneity, so the claimed adaptive weighting may reduce to a static scheme that still suffers the original gradient pathology. A concrete test (e.g., comparison of learned variances inside vs. outside the shock) is needed to substantiate the balancing mechanism.
Validation section (benchmarks): the abstract asserts 'orders of magnitude improvement' without reporting specific L1/L2 error values, shock-capturing metrics, or ablation results against the cited baseline weighting schemes. Without these quantitative anchors the load-bearing claim cannot be assessed independently of the training curves.

minor comments (1)

Notation: the distinction between the spatial mask and the uncertainty-modulated loss terms should be clarified with an explicit equation for the composite loss.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and insightful comments, which help clarify the presentation of our method. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

Referee: Abstract and method description: the central claim of dynamic balancing rests on a single learnable variance parameter per task (homoscedastic aleatoric uncertainty) plus a gradient-based spatial mask. In hyperbolic systems the residual magnitude and its statistics jump by orders of magnitude across shocks; a spatially constant variance cannot capture this heterogeneity, so the claimed adaptive weighting may reduce to a static scheme that still suffers the original gradient pathology. A concrete test (e.g., comparison of learned variances inside vs. outside the shock) is needed to substantiate the balancing mechanism.

Authors: The gradient-based spatial mask is explicitly constructed to modulate the loss contributions according to local gradient magnitude, thereby introducing spatial heterogeneity in the effective weighting even though the per-task variance parameters remain homoscedastic. This design choice directly targets the pathology at discontinuities. To make the mechanism fully transparent, we will add a supplementary figure or table comparing the effective (masked) weights and residual statistics inside versus outside the shock in the revised manuscript. revision: yes
Referee: Validation section (benchmarks): the abstract asserts 'orders of magnitude improvement' without reporting specific L1/L2 error values, shock-capturing metrics, or ablation results against the cited baseline weighting schemes. Without these quantitative anchors the load-bearing claim cannot be assessed independently of the training curves.

Authors: We agree that the abstract would benefit from explicit quantitative anchors. The full manuscript already contains L1/L2 error tables, shock-position errors, and direct comparisons against gradient-based weighting baselines in Section 4. We will revise the abstract to include the key numerical improvements (e.g., specific orders-of-magnitude reductions in L2 error for the Sod and Shu-Osher cases) while respecting length constraints. revision: yes

Circularity Check

0 steps flagged

No derivation chain; empirical method proposal with no self-referential reduction

full rationale

The manuscript introduces UM-PINN as a new probabilistic framework that reinterprets PINN training via homoscedastic aleatoric uncertainty, a gradient-based spatial mask, and learnable variances, then reports experimental gains on standard benchmarks. No equations, theorems, or first-principles derivations are supplied that reduce by construction to the method's own fitted parameters or inputs. Validation is external (Sod tube, Shu-Osher, 2D Riemann) rather than a closed self-referential loop. No self-citation load-bearing steps or ansatz smuggling appear in the provided text. This is the normal non-circular case for an engineering method paper.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The abstract relies on the assumption that PINN training can be recast as homoscedastic uncertainty-weighted multi-task learning and that a gradient mask plus Sobol sampling will stabilize it; no free parameters are numerically specified and no new physical entities are introduced.

free parameters (1)

learnable variance parameters
Introduced to dynamically balance PDE residuals and initial conditions; values are learned during training but not reported.

axioms (2)

domain assumption PINN training difficulties at shocks arise primarily from gradient pathology due to extreme stiffness at discontinuities.
Stated as the main failure mode that the new method addresses.
domain assumption Quasi-Monte Carlo Sobol sampling further stabilizes the framework.
Invoked without justification or comparison to standard sampling.

pith-pipeline@v0.9.0 · 5738 in / 1508 out tokens · 35084 ms · 2026-05-25T06:21:21.026737+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 25 canonical work pages · 1 internal anchor

[1]

R. J. LeVeque, Finite Volume Methods for Hyperbolic Problems, Vol. 31, Cambridge University Press, 2002

work page 2002
[2]

E. F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practical Introduction, Springer Science & Business Media, 2013

work page 2013
[3]

Q. Wang, J. S. Hesthaven, Reduced order modeling of the navier- stokes equations, SIAM Journal on Scientiﬁc Computing 39 (1) (2017) A22–A51

work page 2017
[4]

Karni, Hybrid schemes for euler equations, Journal of Computa- tional Physics 124 (2) (1996) 245–262

S. Karni, Hybrid schemes for euler equations, Journal of Computa- tional Physics 124 (2) (1996) 245–262

work page 1996
[5]

Raissi, P

M. Raissi, P. Perdikaris, 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 378 (2019) 686–707. doi:10.1016/j.jcp. 2018.10.045

work page doi:10.1016/j.jcp 2019
[6]

Z. Tao, K. Xu, F. Liu, LSTM-PINN: An hybrid method for prediction of steady-state electrohydrodynamic ﬂow , Journal of Computational Physics 548 (2026) 114586. doi:10.1016/j.jcp.2025.114586. URL http://dx.doi.org/10.1016/j.jcp.2025.114586

work page doi:10.1016/j.jcp.2025.114586 2026
[7]

L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning non- linear operators via deeponet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (3) (2021) 218– 229

work page 2021
[8]

A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial diﬀerential equa- tions, Communications in Computational Physics 28 (5) (2020) 2002– 2041

work page 2020
[9]

S. Cai, Z. Mao, Z. Wang, et al., Physics-informed neural networks for ﬂuid mechanics: A review, Acta Mechanica Sinica 37 (6) (2021) 951–973

work page 2021
[10]

Z. Mao, A. D. Jagtap, G. E. Karniadakis, Physics-informed neural net- works for high-speed ﬂows, Computer Methods in Applied Mechanics and Engineering 360 (2020) 112789

work page 2020
[11]

Raissi, A

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

work page 2020
[12]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, et al., Physics-informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440

work page 2021
[13]

O. Fuks, H. A. Tchelepi, Limitations of physics-informed machine learning for nonlinear conservation laws, Journal of Machine Learn- ing for Modeling and Computing 1 (1) (2020)

work page 2020
[14]

De Ryck, S

T. De Ryck, S. Lanthaler, S. Mishra, On the approximation of func- tions by relu neural networks with application to pinns, arXiv preprint arXiv:2101.02944 (2022)

work page arXiv 2022
[15]

Krishnapriyan, A

A. Krishnapriyan, A. Ghalue, T. Anderson, et al., Characterizing pos- sible failure modes in physics-informed neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 34, 2021

work page 2021
[16]

A. D. Jagtap, K. Kawaguchi, G. E. Karniadakis, Adaptive activation functions accelerate convergence in deep and physics-informed neural networks, Journal of Computational Physics 404 (2020) 109136

work page 2020
[17]

L. D. McClenny, U. M. Braga-Neto, Self-adaptive physics-informed neural networks using a soft attention mechanism, Journal of Compu- tational Physics 474 (2023) 111722

work page 2023
[18]

Kendall, Y

A. Kendall, Y. Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018, pp. 7482–7491

work page 2018
[19]

A. F. Psaros, X. Meng, Z. Zou, et al., Uncertainty quantiﬁcation in scientiﬁc machine learning: Methods and applications, Journal of Computational Physics 457 (2022) 111052

work page 2022
[20]

X. Jin, S. Cai, H. Li, G. E. Karniadakis, Nsfnets: Physics-informed neural networks for the navier-stokes equations, Journal of Computa- tional Physics 426 (2021) 109951

work page 2021
[21]

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

work page 2021
[22]

I. M. Sobol, On the distribution of points in a cube and the approxi- mate evaluation of integrals, USSR Computational Mathematics and Mathematical Physics 7 (4) (1967) 86–112

work page 1967
[23]

Kasheﬁ, D

A. Kasheﬁ, D. Rempe, L. J. Guibas, T. Mukerji, A point-cloud deep learning framework for prediction of ﬂuid ﬂow ﬁelds on irregular geometries, Physics of Fluids 33 (2) (2021) 027104

work page 2021
[24]

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

work page internal anchor Pith review Pith/arXiv arXiv 2014
[25]

Z. Chen, B. Liu, Y. Li, et al., Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, in: International Conference on Machine Learning (ICML), 2018, pp. 794–803. D. Zhao et al.: Preprint submitted to Elsevier Page 18 of 18

work page 2018

[1] [1]

R. J. LeVeque, Finite Volume Methods for Hyperbolic Problems, Vol. 31, Cambridge University Press, 2002

work page 2002

[2] [2]

E. F. Toro, Riemann Solvers and Numerical Methods for Fluid Dynamics: A Practical Introduction, Springer Science & Business Media, 2013

work page 2013

[3] [3]

Q. Wang, J. S. Hesthaven, Reduced order modeling of the navier- stokes equations, SIAM Journal on Scientiﬁc Computing 39 (1) (2017) A22–A51

work page 2017

[4] [4]

Karni, Hybrid schemes for euler equations, Journal of Computa- tional Physics 124 (2) (1996) 245–262

S. Karni, Hybrid schemes for euler equations, Journal of Computa- tional Physics 124 (2) (1996) 245–262

work page 1996

[5] [5]

Raissi, P

M. Raissi, P. Perdikaris, 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 378 (2019) 686–707. doi:10.1016/j.jcp. 2018.10.045

work page doi:10.1016/j.jcp 2019

[6] [6]

Z. Tao, K. Xu, F. Liu, LSTM-PINN: An hybrid method for prediction of steady-state electrohydrodynamic ﬂow , Journal of Computational Physics 548 (2026) 114586. doi:10.1016/j.jcp.2025.114586. URL http://dx.doi.org/10.1016/j.jcp.2025.114586

work page doi:10.1016/j.jcp.2025.114586 2026

[7] [7]

L. Lu, P. Jin, G. Pang, Z. Zhang, G. E. Karniadakis, Learning non- linear operators via deeponet based on the universal approximation theorem of operators, Nature Machine Intelligence 3 (3) (2021) 218– 229

work page 2021

[8] [8]

A. D. Jagtap, G. E. Karniadakis, Extended physics-informed neural networks (xpinns): A generalized space-time domain decomposition based deep learning framework for nonlinear partial diﬀerential equa- tions, Communications in Computational Physics 28 (5) (2020) 2002– 2041

work page 2020

[9] [9]

S. Cai, Z. Mao, Z. Wang, et al., Physics-informed neural networks for ﬂuid mechanics: A review, Acta Mechanica Sinica 37 (6) (2021) 951–973

work page 2021

[10] [10]

Z. Mao, A. D. Jagtap, G. E. Karniadakis, Physics-informed neural net- works for high-speed ﬂows, Computer Methods in Applied Mechanics and Engineering 360 (2020) 112789

work page 2020

[11] [11]

Raissi, A

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

work page 2020

[12] [12]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, et al., Physics-informed machine learning, Nature Reviews Physics 3 (6) (2021) 422–440

work page 2021

[13] [13]

O. Fuks, H. A. Tchelepi, Limitations of physics-informed machine learning for nonlinear conservation laws, Journal of Machine Learn- ing for Modeling and Computing 1 (1) (2020)

work page 2020

[14] [14]

De Ryck, S

T. De Ryck, S. Lanthaler, S. Mishra, On the approximation of func- tions by relu neural networks with application to pinns, arXiv preprint arXiv:2101.02944 (2022)

work page arXiv 2022

[15] [15]

Krishnapriyan, A

A. Krishnapriyan, A. Ghalue, T. Anderson, et al., Characterizing pos- sible failure modes in physics-informed neural networks, in: Advances in Neural Information Processing Systems (NeurIPS), Vol. 34, 2021

work page 2021

[16] [16]

A. D. Jagtap, K. Kawaguchi, G. E. Karniadakis, Adaptive activation functions accelerate convergence in deep and physics-informed neural networks, Journal of Computational Physics 404 (2020) 109136

work page 2020

[17] [17]

L. D. McClenny, U. M. Braga-Neto, Self-adaptive physics-informed neural networks using a soft attention mechanism, Journal of Compu- tational Physics 474 (2023) 111722

work page 2023

[18] [18]

Kendall, Y

A. Kendall, Y. Gal, R. Cipolla, Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, in: Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR), 2018, pp. 7482–7491

work page 2018

[19] [19]

A. F. Psaros, X. Meng, Z. Zou, et al., Uncertainty quantiﬁcation in scientiﬁc machine learning: Methods and applications, Journal of Computational Physics 457 (2022) 111052

work page 2022

[20] [20]

X. Jin, S. Cai, H. Li, G. E. Karniadakis, Nsfnets: Physics-informed neural networks for the navier-stokes equations, Journal of Computa- tional Physics 426 (2021) 109951

work page 2021

[21] [21]

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

work page 2021

[22] [22]

I. M. Sobol, On the distribution of points in a cube and the approxi- mate evaluation of integrals, USSR Computational Mathematics and Mathematical Physics 7 (4) (1967) 86–112

work page 1967

[23] [23]

Kasheﬁ, D

A. Kasheﬁ, D. Rempe, L. J. Guibas, T. Mukerji, A point-cloud deep learning framework for prediction of ﬂuid ﬂow ﬁelds on irregular geometries, Physics of Fluids 33 (2) (2021) 027104

work page 2021

[24] [24]

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

work page internal anchor Pith review Pith/arXiv arXiv 2014

[25] [25]

Z. Chen, B. Liu, Y. Li, et al., Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, in: International Conference on Machine Learning (ICML), 2018, pp. 794–803. D. Zhao et al.: Preprint submitted to Elsevier Page 18 of 18

work page 2018