Week 46 10 papers Coverage: 2026-11-05 → 2026-11-10

Week 46 · From Running to Trustworthy: Operator Certification, Evaluation Standards and a 14TB Turbulence Benchmark

Half of this week is about making physics AI trustworthy rather than merely runnable. Interval bound propagation and CROWN are lifted into function space so that neural operators can be certified over the entire continuous domain; PINNeval splits PINN training into four components and demands that even the components not under study use competitive implementations; and a 14 TB turbulent cylinder-wake benchmark exposes temporal instability and loss of small-scale energy in autoregressive rollouts. The other half is fundamentals: coordinate transformations to flatten PINN gradient distributions, multifidelity linear regression for reduced operators under data scarcity, and Wasserstein-path data synthesis that cuts solver-free training error by an order of magnitude.

Highlights

  • Neural operators can be certified over a continuous domain for the first time: interval bound propagation and CROWN are formulated in function space, propagating bound functions layer by layer and extended to higher-order derivatives so PDE residuals and solution error bounds can be certified - giving tight, domain-wide guarantees for a 572.5-million-parameter spherical FNO used for global weather forecasting (2611.09038)
  • PINN evaluation is redefined: PINNeval treats domain-importance adaptation, loss balancing, network architecture and optimization scheme as one integrated system, requires competitive implementations even for the components not under study, and ships the JAX package PINNpipe with 18 methods and 9 benchmark problems (accepted at NeurIPS 2026, Evaluations and Datasets track, 2611.08559)
  • 3D turbulent surrogates finally get a hard benchmark: TurtLES covers shear-layer transition cylinder wakes with a 14 TB dataset (400 trajectories, 400 time steps each, 3-11 million points per frame) and identifies three autoregressive rollout failure modes - temporal instability, loss of small-scale energy, and difficulty on large irregular domains (2611.09567)
  • Solver-free operator learning drops an order of magnitude in error: WaSP synthesizes training pairs stage by stage along intermediate distributions that converge to the target distribution, with a stepwise error-propagation bound, cutting solution error by 3.9x to 9.5x versus existing solver-free methods on Allen-Cahn, Kuramoto-Sivashinsky and Ginzburg-Landau (2611.07646)
  • The cause of imbalanced PINN multi-task optimization is quantified: a first-order slow-manifold approximation attributes the step-to-step redistribution of task contributions to curvature-related alignment disparities and coordinatewise differences of preconditioned task gradients, and GMAdam beats AdaTask at comparable cost on high-frequency Poisson, Allen-Cahn and lid-driven cavity flows at Re=100 and 400 (2611.08910)

In this issue

1

Equidistribution of Solution Gradients in Physics-Informed Neural Networks via Coordinate Transformations

Authors Tianchi Yu, Ivan Oseledets
Why read it

PINNs struggle to approximate solutions with sharp gradients because of optimization difficulty and spectral bias. Drawing on the gradient-equidistribution principle from grid generation, this work adds a learnable coordinate transformation that maps the physical domain into a computational domain where the solution becomes approximately affine, and lets the network learn that easier function. Two constructions are given: an explicit one based on Majorants and an implicit one based on invertible neural networks, the latter usable as a plug-in module alongside stronger PINN models to improve accuracy across challenging PDE benchmarks.

arXiv · 2611.09053 ↗ physics-informedpinncoordinate-transformationspectral-bias
2

Function-Space Certification for Neural Operators

Authors Ryan Hsiang, Robert Joseph George, Anima Anandkumar 45 pages, 6 figures
Why read it

Standard neural network verification is designed for vector-to-vector maps, so applied to neural operators it can only certify a fixed grid resolution and offers no guarantee across another resolution or over the whole domain. This work formulates interval bound propagation and CROWN in function space, propagating upper and lower bound functions through each layer using rigorous numerical integration, and extends bound propagation to higher-order derivatives so that PDE residuals and solution error bounds can be certified. Across several operator architectures and scientific problems, including a 572.5-million-parameter spherical Fourier neural operator for global weather forecasting, the resulting guarantees are tight and domain-wide.

arXiv · 2611.09038 ↗ neural-operatorverificationcertificationrobustness
3

Multifidelity Operator Inference for Non-intrusive Model Reduction under Data Scarcity

Authors Vivian Zhang, Tomoki Koike, Atticus Rex et al. 21 pages, AIAA SciTech Forum 2027
Why read it

Operator inference fits reduced operators to state and time-derivative data for non-intrusive surrogate modelling, but high-fidelity training data is scarce under a limited compute budget, and models trained on such data exhibit high variance - a critical liability in high-risk engineering design. Multifidelity operator inference reduces that variance by supplementing scarce high-fidelity snapshots with abundant low-fidelity data drawn from coarser discretizations of the same problem, handling all fidelity levels in a single multifidelity linear regression framework that estimates the reduced operators. On model PDE problems it predicts solution fields more accurately than high-fidelity-only operator inference at the same computational budget.

arXiv · 2611.13624 ↗ reduced-orderoperator-inferencemultifidelitysurrogate-model
4

WaSP: Wasserstein Data Synthesis Pipeline for Solver-Free Operator Learning

Authors Ryan Liu, Chuwei Wang, Ricardo Baptista et al.
Why read it

Two solver-free alternatives to training neural operators both have drawbacks: physics-informed learning needs a strong residual norm to control solution error, but a stronger norm raises mini-batch variance and hinders optimization, while manufactured data can be mutually singular with the target distribution. WaSP trains on manufactured input-solution pairs while gradually mitigating the shift, synthesizing new pairs by applying the learned operator along a path of intermediate distributions that converges to the target input distribution, where each stage lets the previous surrogate generate solutions and the exact PDE operator derive the corresponding inputs. The authors prove a stepwise error-propagation bound for the shift introduced during synthesis, and report solution errors 3.9x to 9.5x lower than existing solver-free methods on Allen-Cahn, Kuramoto-Sivashinsky and Ginzburg-Landau, with better reproduction of statistical observables for two chaotic systems.

arXiv · 2611.07646 ↗ neural-operatorsolver-freedata-synthesisdistribution-shift
5

A scaled TW-PINN framework for soliton solutions of the Korteweg--de Vries equation

Authors Seungwan Han, Jiaxi Gu, Daniel Olmos-Liceaga et al.
Why read it

Soliton solutions of the Korteweg-de Vries equation remain challenging for PINNs because of localized wave structures, high-order spatial derivatives, nonlinear interaction between solitons and long-time prediction. This scaling framework transforms the governing equation into a normalized form, approximates the solution with a PINN solver, and reconstructs it in the original variables through inverse scaling. The one-soliton solver adds a traveling-wave coordinate to capture propagation, while the two-soliton case uses a two-step training procedure that first learns individual soliton profiles and propagation speeds, then builds traveling-wave coordinates from those speeds to approximate the full solution including nonlinear interactions. A soliton-informed sampling scheme places extra collocation points around soliton trajectories. Results show accurate single-soliton propagation, resolved two-soliton interactions and preserved soliton profiles during long-time extrapolation, with an additional three-soliton experiment suggesting the two-step procedure extends to multi-soliton solutions.

arXiv · 2611.09339 ↗ physics-informedpinnsolitontraveling-wave
6

LithoSAGE: Spatially Adaptive and Geometrically Equivariant Neural Operators for Computational Lithography

Authors Ke Jin, Zhenli Xu, Zilin You
Why read it

Operator learning can replace costly lithography simulation with direct model inference, but existing architectures are not designed to provide geometric equivariance and adaptability to layout windows with different physical extents at the same time, which limits generalization. LithoSAGE works inside a Fourier neural operator and expresses spectral weights as continuous functions of normalized radial frequency through a learnable sum of Gaussians, preserving the geometric properties of the lithography system while incorporating rotation and translation equivariance, spatial interaction decay and direct evaluation on different layout windows. Experiments predict aerial images from mask layouts and resist images from aerial images, demonstrating geometric equivariance and spatial adaptivity, and compare prediction error and parameter efficiency with other architectures.

arXiv · 2611.08001 ↗ neural-operatorequivariancelithographysemiconductor
7

EDP-PI-DOSnet: An Energy Dissipation Preserving Neural Solver for Two-Phase Flow Problems

Authors Jizu Huang, Yue Qian, Tao Zhou
Why read it

Preserving the energy dissipation properties of gradient flow equations matters for the stability and reliability of numerical approximations, particularly in long-time simulation. This paper proposes energy-dissipation-preserving physics-informed deep operator splitting networks that incorporate classical structure-preserving numerical schemes into the architecture, imposing constraints on the convolution kernels so that the learned discrete operators are symmetric and negative semidefinite. Several network formulations are derived from the Crank-Nicolson, scalar auxiliary variable Crank-Nicolson and relaxed discrete variational derivative schemes, together with a physics-informed loss based on the residuals of the governing equations. Numerical experiments show predicted solutions in good agreement with reference solutions while satisfying the corresponding discrete energy dissipation laws, stable behaviour during long-time recursive prediction, and potential applicability to two-phase flow problems.

arXiv · 2611.09430 ↗ physics-informedstructure-preservingtwo-phase-flowdeep-operator
8

Task-Contribution Dynamics of Adam in Physics-Informed Neural Networks: Slow-Manifold Analysis and Gradient Modulation

Authors Enzhan Zheng, Xiaoliang Song
Why read it

PINNs often suffer from imbalanced optimization of competing loss components, particularly when approximating solutions with high-frequency or multiscale features, yet how each task contribution evolves into Adam parameter updates was insufficiently understood. This paper introduces task-wise indicators that identify three distinct training regimes - smooth variation, intermittent spikes and persistent oscillations - and, in the early smooth regime, combines discrete-time analysis with a continuous-time singular perturbation framework under full-batch training and the parameter regime (1-beta2)/(1-beta1) much smaller than 1. A first-order approximation of the locally attracting slow manifold reveals two factors governing the step-to-step redistribution of task contributions: curvature-related alignment disparities and differences in the coordinatewise distributions of preconditioned task gradients, both of which involve the shared second-moment estimate. Inspired by AdaTask, the authors propose Gradient-Modulated Adam, which modulates incoming task gradients using task-specific second-moment estimates; experiments on high-frequency Poisson and wave equations, the Allen-Cahn equation and two-dimensional lid-driven cavity flows at Re=100 and 400 show consistent improvements over AdaTask at comparable computational cost.

arXiv · 2611.08910 ↗ physics-informedpinnoptimizationmulti-task
9

TurtLES: A Large-Scale Benchmark for Turbulent 3D Neural PDE Surrogates

Authors Jarno Platenburg, Armand Kassaï Koupaï, Paola Cinnella et al.
Why read it

Machine-learning surrogates are increasingly used to accelerate computational fluid dynamics, yet progress is limited by the lack of benchmarks capturing realistic, time-dependent turbulent flows. TurtLES is a 14 TB dataset of three-dimensional turbulent cylinder wakes in the shear-layer transition regime, computed via high-fidelity implicit large-eddy simulations: 400 trajectories of 400 time steps each, with 3 to 11 million points per frame, spanning variations in geometry, Reynolds number and angle of attack. Unlike existing datasets it combines an active energy cascade driven by vortex stretching, dense irregular meshes and long-horizon temporal dynamics. The authors benchmark state-of-the-art neural operators on autoregressive rollouts and identify key failure modes including temporal instability, loss of small-scale energy, and challenges in learning on large irregular domains.

arXiv · 2611.09567 ↗ neural-operatorbenchmarkturbulencedataset
10

PINNeval: A Comprehensive Evaluation Standard for Physics-Informed Neural Networks

Authors Kevin von Bargen, Hamid Ebrahimy, Tim Römer Accepted at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), Track on Evaluations and Datasets. 54 pages. Code: https://github.com/KevinvonBargen/pinnpipe
Why read it

The PINN literature has produced a rapidly expanding catalogue of methods, yet many existing benchmarks study these components in isolation and set the surrounding, not considered components of the pipeline to a minimal default that no longer represents current best practice, so conclusions drawn under such conditions may not transfer to competitive pipelines. This work proposes an evaluation protocol that treats the four components - domain-importance adaptation, loss balancing, neural network architecture and optimization scheme - as an integrated system, crucially requires competitive implementations of the components not under investigation, specifies the metrics and reporting format, and is accompanied by structured documentation ensuring traceability of evaluations across studies. Based on this protocol the authors release PINNpipe, a JAX package containing 18 methods from the current literature and nine efficiently implemented benchmark problems, and illustrate the protocol against established benchmark studies with results that challenge common assumptions.

arXiv · 2611.08559 ↗ physics-informedpinnbenchmarkevaluationjax

Overview

Half of this week is about making physics AI trustworthy rather than merely runnable. Interval bound propagation and CROWN are lifted into function space so that neural operators can be certified over the entire continuous domain; PINNeval splits PINN training into four components and demands that even the components not under study use competitive implementations; and a 14 TB turbulent cylinder-wake benchmark exposes temporal instability and loss of small-scale energy in autoregressive rollouts. The other half is fundamentals: coordinate transformations to flatten PINN gradient distributions, multifidelity linear regression for reduced operators under data scarcity, and Wasserstein-path data synthesis that cuts solver-free training error by an order of magnitude.

1. Equidistribution of Solution Gradients in Physics-Informed Neural Networks via Coordinate Transformations

PINNs struggle to approximate solutions with sharp gradients because of optimization difficulty and spectral bias. Drawing on the gradient-equidistribution principle from grid generation, this work adds a learnable coordinate transformation that maps the physical domain into a computational domain where the solution becomes approximately affine, and lets the network learn that easier function. Two constructions are given: an explicit one based on Majorants and an implicit one based on invertible neural networks, the latter usable as a plug-in module alongside stronger PINN models to improve accuracy across challenging PDE benchmarks.

2. Function-Space Certification for Neural Operators

Standard neural network verification is designed for vector-to-vector maps, so applied to neural operators it can only certify a fixed grid resolution and offers no guarantee across another resolution or over the whole domain. This work formulates interval bound propagation and CROWN in function space, propagating upper and lower bound functions through each layer using rigorous numerical integration, and extends bound propagation to higher-order derivatives so that PDE residuals and solution error bounds can be certified. Across several operator architectures and scientific problems, including a 572.5-million-parameter spherical Fourier neural operator for global weather forecasting, the resulting guarantees are tight and domain-wide.

3. Multifidelity Operator Inference for Non-intrusive Model Reduction under Data Scarcity

Operator inference fits reduced operators to state and time-derivative data for non-intrusive surrogate modelling, but high-fidelity training data is scarce under a limited compute budget, and models trained on such data exhibit high variance - a critical liability in high-risk engineering design. Multifidelity operator inference reduces that variance by supplementing scarce high-fidelity snapshots with abundant low-fidelity data drawn from coarser discretizations of the same problem, handling all fidelity levels in a single multifidelity linear regression framework that estimates the reduced operators. On model PDE problems it predicts solution fields more accurately than high-fidelity-only operator inference at the same computational budget.

4. WaSP: Wasserstein Data Synthesis Pipeline for Solver-Free Operator Learning

Two solver-free alternatives to training neural operators both have drawbacks: physics-informed learning needs a strong residual norm to control solution error, but a stronger norm raises mini-batch variance and hinders optimization, while manufactured data can be mutually singular with the target distribution. WaSP trains on manufactured input-solution pairs while gradually mitigating the shift, synthesizing new pairs by applying the learned operator along a path of intermediate distributions that converges to the target input distribution, where each stage lets the previous surrogate generate solutions and the exact PDE operator derive the corresponding inputs. The authors prove a stepwise error-propagation bound for the shift introduced during synthesis, and report solution errors 3.9x to 9.5x lower than existing solver-free methods on Allen-Cahn, Kuramoto-Sivashinsky and Ginzburg-Landau, with better reproduction of statistical observables for two chaotic systems.

5. A scaled TW-PINN framework for soliton solutions of the Korteweg--de Vries equation

Soliton solutions of the Korteweg-de Vries equation remain challenging for PINNs because of localized wave structures, high-order spatial derivatives, nonlinear interaction between solitons and long-time prediction. This scaling framework transforms the governing equation into a normalized form, approximates the solution with a PINN solver, and reconstructs it in the original variables through inverse scaling. The one-soliton solver adds a traveling-wave coordinate to capture propagation, while the two-soliton case uses a two-step training procedure that first learns individual soliton profiles and propagation speeds, then builds traveling-wave coordinates from those speeds to approximate the full solution including nonlinear interactions. A soliton-informed sampling scheme places extra collocation points around soliton trajectories. Results show accurate single-soliton propagation, resolved two-soliton interactions and preserved soliton profiles during long-time extrapolation, with an additional three-soliton experiment suggesting the two-step procedure extends to multi-soliton solutions.

6. LithoSAGE: Spatially Adaptive and Geometrically Equivariant Neural Operators for Computational Lithography

Operator learning can replace costly lithography simulation with direct model inference, but existing architectures are not designed to provide geometric equivariance and adaptability to layout windows with different physical extents at the same time, which limits generalization. LithoSAGE works inside a Fourier neural operator and expresses spectral weights as continuous functions of normalized radial frequency through a learnable sum of Gaussians, preserving the geometric properties of the lithography system while incorporating rotation and translation equivariance, spatial interaction decay and direct evaluation on different layout windows. Experiments predict aerial images from mask layouts and resist images from aerial images, demonstrating geometric equivariance and spatial adaptivity, and compare prediction error and parameter efficiency with other architectures.

7. EDP-PI-DOSnet: An Energy Dissipation Preserving Neural Solver for Two-Phase Flow Problems

Preserving the energy dissipation properties of gradient flow equations matters for the stability and reliability of numerical approximations, particularly in long-time simulation. This paper proposes energy-dissipation-preserving physics-informed deep operator splitting networks that incorporate classical structure-preserving numerical schemes into the architecture, imposing constraints on the convolution kernels so that the learned discrete operators are symmetric and negative semidefinite. Several network formulations are derived from the Crank-Nicolson, scalar auxiliary variable Crank-Nicolson and relaxed discrete variational derivative schemes, together with a physics-informed loss based on the residuals of the governing equations. Numerical experiments show predicted solutions in good agreement with reference solutions while satisfying the corresponding discrete energy dissipation laws, stable behaviour during long-time recursive prediction, and potential applicability to two-phase flow problems.

8. Task-Contribution Dynamics of Adam in Physics-Informed Neural Networks: Slow-Manifold Analysis and Gradient Modulation

PINNs often suffer from imbalanced optimization of competing loss components, particularly when approximating solutions with high-frequency or multiscale features, yet how each task contribution evolves into Adam parameter updates was insufficiently understood. This paper introduces task-wise indicators that identify three distinct training regimes - smooth variation, intermittent spikes and persistent oscillations - and, in the early smooth regime, combines discrete-time analysis with a continuous-time singular perturbation framework under full-batch training and the parameter regime (1-beta2)/(1-beta1) much smaller than 1. A first-order approximation of the locally attracting slow manifold reveals two factors governing the step-to-step redistribution of task contributions: curvature-related alignment disparities and differences in the coordinatewise distributions of preconditioned task gradients, both of which involve the shared second-moment estimate. Inspired by AdaTask, the authors propose Gradient-Modulated Adam, which modulates incoming task gradients using task-specific second-moment estimates; experiments on high-frequency Poisson and wave equations, the Allen-Cahn equation and two-dimensional lid-driven cavity flows at Re=100 and 400 show consistent improvements over AdaTask at comparable computational cost.

9. TurtLES: A Large-Scale Benchmark for Turbulent 3D Neural PDE Surrogates

Machine-learning surrogates are increasingly used to accelerate computational fluid dynamics, yet progress is limited by the lack of benchmarks capturing realistic, time-dependent turbulent flows. TurtLES is a 14 TB dataset of three-dimensional turbulent cylinder wakes in the shear-layer transition regime, computed via high-fidelity implicit large-eddy simulations: 400 trajectories of 400 time steps each, with 3 to 11 million points per frame, spanning variations in geometry, Reynolds number and angle of attack. Unlike existing datasets it combines an active energy cascade driven by vortex stretching, dense irregular meshes and long-horizon temporal dynamics. The authors benchmark state-of-the-art neural operators on autoregressive rollouts and identify key failure modes including temporal instability, loss of small-scale energy, and challenges in learning on large irregular domains.

10. PINNeval: A Comprehensive Evaluation Standard for Physics-Informed Neural Networks

The PINN literature has produced a rapidly expanding catalogue of methods, yet many existing benchmarks study these components in isolation and set the surrounding, not considered components of the pipeline to a minimal default that no longer represents current best practice, so conclusions drawn under such conditions may not transfer to competitive pipelines. This work proposes an evaluation protocol that treats the four components - domain-importance adaptation, loss balancing, neural network architecture and optimization scheme - as an integrated system, crucially requires competitive implementations of the components not under investigation, specifies the metrics and reporting format, and is accompanied by structured documentation ensuring traceability of evaluations across studies. Based on this protocol the authors release PINNpipe, a JAX package containing 18 methods from the current literature and nine efficiently implemented benchmark problems, and illustrate the protocol against established benchmark studies with results that challenge common assumptions.