REVIEW 9 cited by
Characterizing possible failure modes in physics-informed neural networks
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Recent work in scientific machine learning has developed so-called physics-informed neural network (PINN) models. The typical approach is to incorporate physical domain knowledge as soft constraints on an empirical loss function and use existing machine learning methodologies to train the model. We demonstrate that, while existing PINN methodologies can learn good models for relatively trivial problems, they can easily fail to learn relevant physical phenomena for even slightly more complex problems. In particular, we analyze several distinct situations of widespread physical interest, including learning differential equations with convection, reaction, and diffusion operators. We provide evidence that the soft regularization in PINNs, which involves PDE-based differential operators, can introduce a number of subtle problems, including making the problem more ill-conditioned. Importantly, we show that these possible failure modes are not due to the lack of expressivity in the NN architecture, but that the PINN's setup makes the loss landscape very hard to optimize. We then describe two promising solutions to address these failure modes. The first approach is to use curriculum regularization, where the PINN's loss term starts from a simple PDE regularization, and becomes progressively more complex as the NN gets trained. The second approach is to pose the problem as a sequence-to-sequence learning task, rather than learning to predict the entire space-time at once. Extensive testing shows that we can achieve up to 1-2 orders of magnitude lower error with these methods as compared to regular PINN training.
Forward citations
Cited by 9 Pith papers
-
Evolution-Level Quantum Optimal Control of Single-Qubit Gates with Physics-Informed Neural Networks
Single-qubit gates are optimized as complete differentiable evolutions, recovering time-optimal pulse structure and enabling local refinement of a geometric Z gate.
-
Diagnosing Failure Modes of Neural Operators Across Diverse PDE Families
A single-architecture empirical study reports degradation factors up to ~18 for neural PDE solvers under distribution shift, but its headline 'architecture-dependent failure patterns' claim is not backed by the experi...
-
LieSolver: PDE-Constrained Learning for IBVPs via Lie Symmetries
LieSolver fits initial/boundary data with linear combinations of Lie-symmetry-generated base solutions, enforcing linear homogeneous PDEs exactly by construction.
-
LIGO-PINN: Learned Initialization via Gated Optimization to Alleviate Convergence Failures in Physics Informed Neural Networks
Meta-learning on easy PDE tasks plus a layer-wise gating schedule reduces extrapolation error by about 91% relative to six PINN baselines on hard convection, Helmholtz, and Navier-Stokes benchmarks.
-
Multi-Head Neural Operator for Modelling Interfacial Dynamics
The Multi-Head Neural Operator predicts full phase-field trajectories in a single forward pass using time-specific projection heads with temporal connections, and outperforms FNO-2d and FNO-3d on five benchmark equations.
-
Operator-based machine learning framework for generalizable prediction of unsteady treatment dynamics in stormwater infrastructure
A composite operator neural network predicts 3D unsteady flow and particle concentrations in a stormwater separator, matching CFD closely on most held-out storm events.
-
Drag modelling for flows through assemblies of spherical particles with machine learning: A comparison of approaches
Applying genetic programming to outputs of a graph neural network yields compact symbolic drag-variation formulas at Reynolds numbers up to 280, though with lower accuracy than the network.
-
Breaking the Precision Ceiling in Physics-Informed Neural Networks: A Hybrid Fourier-Neural Architecture for Ultra-High Accuracy
A Fourier-neural PINN achieves 1.94e-7 L2 error on a beam equation, but the 'ultra-precision' is probably due to the solution being exactly representable by the chosen Fourier modes.
-
Towards Digital Twins for Optimal Radioembolization
A review proposing a liver radioembolization digital twin that combines CFD with physics-informed neural networks to plan microsphere delivery.
Discussion (0). Sign in to comment.