AI models designed to mimic the laws of physics are only pretending to understand the world by sticking to the simplest possible paths.
April 25, 2026
Original Paper
Predictivity and Utility of Neural Surrogates of Multiscale PDEs
arXiv · 2604.20061
The Takeaway
Modern emulators for partial differential equations succeed by operating on low-dimensional manifolds while ignoring high-frequency details. This research shows these models fundamentally cannot recover the complex information lost during the simplification of physical data. Many experts believed that AI could eventually replace classical solvers for chaotic tasks like weather forecasting or fluid dynamics. In reality, these models crash when faced with the high-energy turbulence that defines real-world physics. Scientists must now stop viewing AI as a total replacement and start using it only for the simplified parts of the simulation.
From the abstract
Scientific machine learning is increasingly being spoken of as universal emulators for classical numerical solvers for multi-scale partial differential equations, but most apparent successes can be explained by facts that also define their limits. Many successful benchmarks live on low-dimensional solution manifolds where any competent reduced model will interpolate well. More fundamentally, neural surrogates systematically under-resolve high-frequency content due to spectral bias, and coarse-gr