Week 43 · Beyond Residuals: Verifiability and Structural Priors in Neural Operators
None of this week's ten papers stops at "smaller residual": they ask whether the solution is right. A viscosity-solution test shows a tiny residual can coexist with an entirely wrong value; reduced-order optimality is formalized in reproducing kernel Hilbert spaces; and the extrapolation limit of neural operators is located in geometric support rather than in architecture.
Highlights
- Small residual, wrong solution: across 30 runs the standard periodic network always converged to a folded solution that is right nowhere
- Reduced-order models gain optimality theory: interpolatory conditions generalized to RKHS, including non-differentiable kernels
- Unstructured meshes with rigorous bounds: the learnable-spline operator reaches 46M nodes and certifies predictions via Bernstein convex hulls
- Extrapolation limits traced to geometric support: active learning and physical constraints work better as fine-tuning objectives
In this issue
Norm-optimal model order reduction in reproducing kernel Hilbert spaces
The classical necessary conditions for norm-optimal reduction lived only in H2 Hardy-Hilbert spaces; this generalizes them to reproducing kernel Hilbert spaces, with orthonormal bases for arbitrary RKHS and an algorithm covering both differentiable and non-differentiable kernels. A bonus: solving PDEs with the optimal kernel basis yields a FEM-like method.
Low Residuals, Wrong Value: Learning the Viscosity Solution of Hamilton-Jacobi-Isaacs Equations with PINNs
First-order Hamilton-Jacobi equations admit almost-everywhere solutions that violate the equation only on measure-zero sets, differ from the viscosity solution by O(1), and look perfectly fine to residual training. This quantifies the trap across 120 runs and adds a structural test that flags wrong solutions without the ground truth.
Synthetic Dataset Generation, Physical Consistency and Active Learning for Shape-Informed Surrogate Modeling of Hemodynamics
Inter-patient geometric variability is the real bottleneck for hemodynamic surrogates. This sidesteps registration with a common reference geometry plus displacement fields, synthesizes physiologically consistent CFD data, and lands a counterintuitive result: active learning and physical consistency both work better as fine-tuning objectives than as regularizers used throughout training.
A mixed leading variable physics-informed neural network for nonsmooth dynamics simulation in contact systems
Stick-slip in frictional contacts is usually handled by first discretizing into complementary subproblems, which costs accuracy and generality. This makes the network solve for acceleration, normal contact force and friction margin simultaneously, with a mixed leading variable that sharpens stick-slip identification; four benchmarks run from 1-D cases to moving contact on an elastic beam.
Learnable Splines for Unstructured Neural Operators
FNO assumes uniform rectangular grids while engineering runs on large, irregular, locally refined meshes. Learnable tensor-product B-splines bridge the two, reaching 26.6M training nodes and 46M at inference on one A100, with rigorous prediction bounds over continuous space on top.
Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization
Reduced-order models usually wait for the high-fidelity solve to converge, which is expensive. This builds local models straight from unconverged primal iterations and rebuilds them as the design moves, giving 8-22x speedups on three PDE-constrained optimization problems and still about 3x on a transonic inverse design with moving shocks.
SBNO : Schrödinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information
When observations are sparse and the physics is missing, diffusion surrogates usually need guidance weights tuned per instance. This learns conditional operators with Schrödinger bridge matching, handles forward and inverse problems in one framework, and reconstructs deterministically via the probability-flow ODE — over 9.7x faster than guided diffusion at 5% observations.
SINO: Scale-Invariant Neural Operator
Closure terms on coarse grids follow scale-invariant physical laws even though they depend on resolution. Encoding that low-rank bias explicitly concentrates over 95% of the variance into two or three modes, yields scaling-law exponents 38 times steeper than FNO, and cuts error by 1.5-38x across five turbulent closure problems.
PI-NOMT: Physics-Informed Neural Optimal Mass Transport for Brain Fluid Dynamics
Most inverse problems reconstruct the observable and leave the transport mechanism black-boxed. Here density, velocity and source are all continuous neural fields, with the governing equations acting as a structural prior on which transport mechanisms are admissible — sub-percent endpoint error on rat DCE-MRI and physically consistent speed scales.
Spherical Physics-informed Neural Operator with Multi-scale Coupling for Meteorological Downscaling
Global-to-regional downscaling fights three things at once: spherical geometric distortion, multi-scale coupling mismatch and missing atmospheric laws. Spherical Laplacian decomposition keeps point-wise correspondence across scales, while a geometry-aware localized spherical integral operator handles continuous kernel transforms — with zero-shot generalization to unseen query locations.
Overview
None of this week's ten papers stops at "smaller residual": they ask whether the solution is right. A viscosity-solution test shows a tiny residual can coexist with an entirely wrong value; reduced-order optimality is formalized in reproducing kernel Hilbert spaces; and the extrapolation limit of neural operators is located in geometric support rather than in architecture.
1. Norm-optimal model order reduction in reproducing kernel Hilbert spaces
The classical necessary conditions for norm-optimal reduction lived only in H2 Hardy-Hilbert spaces; this generalizes them to reproducing kernel Hilbert spaces, with orthonormal bases for arbitrary RKHS and an algorithm covering both differentiable and non-differentiable kernels. A bonus: solving PDEs with the optimal kernel basis yields a FEM-like method.
2. Low Residuals, Wrong Value: Learning the Viscosity Solution of Hamilton-Jacobi-Isaacs Equations with PINNs
First-order Hamilton-Jacobi equations admit almost-everywhere solutions that violate the equation only on measure-zero sets, differ from the viscosity solution by O(1), and look perfectly fine to residual training. This quantifies the trap across 120 runs and adds a structural test that flags wrong solutions without the ground truth.
3. Synthetic Dataset Generation, Physical Consistency and Active Learning for Shape-Informed Surrogate Modeling of Hemodynamics
Inter-patient geometric variability is the real bottleneck for hemodynamic surrogates. This sidesteps registration with a common reference geometry plus displacement fields, synthesizes physiologically consistent CFD data, and lands a counterintuitive result: active learning and physical consistency both work better as fine-tuning objectives than as regularizers used throughout training.
4. A mixed leading variable physics-informed neural network for nonsmooth dynamics simulation in contact systems
Stick-slip in frictional contacts is usually handled by first discretizing into complementary subproblems, which costs accuracy and generality. This makes the network solve for acceleration, normal contact force and friction margin simultaneously, with a mixed leading variable that sharpens stick-slip identification; four benchmarks run from 1-D cases to moving contact on an elastic beam.
5. Learnable Splines for Unstructured Neural Operators
FNO assumes uniform rectangular grids while engineering runs on large, irregular, locally refined meshes. Learnable tensor-product B-splines bridge the two, reaching 26.6M training nodes and 46M at inference on one A100, with rigorous prediction bounds over continuous space on top.
6. Online reduced-order modeling from primal solver trajectories for PDE-constrained optimization
Reduced-order models usually wait for the high-fidelity solve to converge, which is expensive. This builds local models straight from unconverged primal iterations and rebuilds them as the design moves, giving 8-22x speedups on three PDE-constrained optimization problems and still about 3x on a transonic inverse design with moving shocks.
7. SBNO : Schrödinger Bridge Neural Operator for the Forward and Inverse Problems under Missing Information
When observations are sparse and the physics is missing, diffusion surrogates usually need guidance weights tuned per instance. This learns conditional operators with Schrödinger bridge matching, handles forward and inverse problems in one framework, and reconstructs deterministically via the probability-flow ODE — over 9.7x faster than guided diffusion at 5% observations.
8. SINO: Scale-Invariant Neural Operator
Closure terms on coarse grids follow scale-invariant physical laws even though they depend on resolution. Encoding that low-rank bias explicitly concentrates over 95% of the variance into two or three modes, yields scaling-law exponents 38 times steeper than FNO, and cuts error by 1.5-38x across five turbulent closure problems.
9. PI-NOMT: Physics-Informed Neural Optimal Mass Transport for Brain Fluid Dynamics
Most inverse problems reconstruct the observable and leave the transport mechanism black-boxed. Here density, velocity and source are all continuous neural fields, with the governing equations acting as a structural prior on which transport mechanisms are admissible — sub-percent endpoint error on rat DCE-MRI and physically consistent speed scales.
10. Spherical Physics-informed Neural Operator with Multi-scale Coupling for Meteorological Downscaling
Global-to-regional downscaling fights three things at once: spherical geometric distortion, multi-scale coupling mismatch and missing atmospheric laws. Spherical Laplacian decomposition keeps point-wise correspondence across scales, while a geometry-aware localized spherical integral operator handles continuous kernel transforms — with zero-shot generalization to unseen query locations.