When PINNs Go Wrong: Pseudo-Time Stepping Against Spurious Solutions

Item: When PINNs Go Wrong: Pseudo-Time Stepping Against Spurious Solutions
Rating: 4
Author: PostSingularity Institute

by Sifan Wang

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQNNeural networks and fuzzy systems

🎙 Sifan Wang 👥 4K 📅 June 13, 2026 ⏱ 68 min 👁 449 📄 expert opinion 🧭 2026-08-15

Available in: English (current) Français

Keywords

PINNspurious solutionspseudo-time steppingPDEcollocation points

Summary

Sifan Wang presents a seminar on the failure modes of Physics-Informed Neural Networks (PINNs) and proposes a pseudo-time stepping method to address them. He begins by introducing the PINN framework for solving forward PDE problems, highlighting common training pathologies such as loss imbalance, spectral bias, and ill-conditioned loss landscapes. He then demonstrates that even with existing techniques, PINNs can converge to spurious, non-physical solutions, especially for challenging problems like linear advection with long time horizons or lid-driven cavity flow at high Reynolds numbers. The core theoretical contribution is a theorem showing that for any finite set of collocation points, there exists a smooth function that satisfies all initial/boundary conditions and achieves zero empirical PINN loss but is trivial after a certain time. This is due to the richness of the neural network function class and the pointwise nature of the residual loss. Wang argues that pseudo-time stepping, a classical technique, helps mitigate this issue not by improving conditioning as previously thought, but by enabling the network to progressively learn the solution from the initial/boundary conditions, especially when combined with random collocation point resampling. He presents empirical results on Burgers’ equation and other benchmarks, showing that pseudo-time stepping with resampling avoids spurious solutions, while fixed collocation points fail. He also discusses the importance of step size selection and introduces an adaptive strategy based on a finite-difference surrogate of the local residual Jacobian. The talk concludes with a summary of findings and implications for reliable physics-informed learning.

245 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides significant value by identifying a fundamental weakness in the PINN loss formulation: the existence of spurious solutions that can be exactly zero empirical loss. This is a novel theoretical insight that goes beyond typical optimization-based explanations. The argumentation is solid, with a clear theorem and proof, and empirical evidence supporting the claims. The speaker effectively contrasts the behavior of pseudo-time stepping with and without resampling, demonstrating that the benefit is not conditioning but rather the progressive learning of the solution. The discussion of step size sensitivity and the proposed adaptive method adds practical value. The presentation is well-structured and the reasoning is transparent, though some claims are based on specific experiments that may not generalize.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor through a formal theorem and proof, and by referencing prior work on PINN pathologies and pseudo-time stepping. The speaker cites specific papers (e.g., TSONN, combustion PINN, P2 PINN) but does not provide URLs in the description. The title accurately reflects the content. The presentation is a seminar talk, so it is not peer-reviewed, but the speaker is a recognized expert. The description includes an abstract that outlines the main contributions. No comments were provided for analysis.

214 words

Title / Content Match

The title accurately reflects the content, which focuses on diagnosing why PINNs fail and proposing a pseudo-time stepping method to mitigate spurious solutions.

Quality & Reliability

8/10

The talk presents a rigorous theoretical analysis of PINN failures, supported by a theorem with proof and empirical demonstrations. The speaker is a recognized researcher in the field. However, the presentation is a seminar talk, not a peer-reviewed publication, and some claims are based on specific experiments.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

When PINNs Go Wrong: Pseudo-Time Stepping Against Spurious Solutions (abstract) — The video description contains the abstract of the talk, which outlines the main contributions.

Concurring Sources

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations — Original PINN paper by Raissi et al., which the talk builds upon.

Contribution & Novelties

The talk provides a novel theoretical explanation for PINN failures: the existence of spurious solutions that achieve zero empirical loss due to the richness of neural networks and finite collocation points. It also clarifies that pseudo-time stepping’s effectiveness is not due to improved conditioning but rather its ability to guide the optimization away from spurious solutions when combined with random resampling. The adaptive step size selection based on a finite-difference surrogate of the local residual Jacobian is a practical contribution.

Pour aller plus loin :

Physics-informed neural networks — Overview of PINNs and their applications.
Pseudo-transient continuation — Classical technique for solving steady-state problems.
Burgers’ equation — Relevant PDE example used in the talk.

114 words

Radar Profile

The radar profile shows high scores in technical level and information quality, reflecting the advanced and rigorous nature of the talk. The lower score in information quantity is due to the focused scope of a single research presentation. Overall, the talk is highly technical and reliable, suitable for experts in the field.

Reliability 8/10