Google DeepMind Discovers New Fluid Dynamics Singularities Using AI

Google DeepMind, in collaboration with researchers from Brown University, New York University, and Stanford University, has introduced a new method for discovering unstable singularities in fluid dynamics. This approach leverages Physics-Informed Neural Networks (PINNs) and high-precision optimization to identify mathematical "blow ups" where quantities like pressure or velocity become infinite, helping mathematicians identify fundamental limitations in the equations governing fluid motion.

Discovery of Unstable Singularities in Fluid Equations

DeepMind's method has enabled the first systematic discovery of new families of unstable singularities across three different fluid equations. The researchers observed a consistent pattern where the speed of the blow up (represented by the parameter lambda, Δ) is correlated with the order of instability (the number of ways a solution can deviate from the blow up). This pattern was specifically visible in the Incompressible Porous Media (IPM) and Boussinesq equations, suggesting that further unstable solutions exist along the same trajectory.

Technical Implementation: High-Precision PINNs

The discovery was made possible by transforming Physics-Informed Neural Networks (PINNs) from general-purpose PDE solvers into specialized discovery tools. The technical framework relies on several key components:

  • Physics-Informed Training: Unlike standard neural networks that rely on large datasets, these models are trained to match the physical laws described by the equations. The network minimizes its "residual"—the degree to which its solution fails to satisfy the physical equations.
  • Second-Order Optimizers: The team incorporated second-order optimizers for training, which allowed the model to reach unprecedented levels of accuracy.
  • Near-Machine Precision: The framework pushes PINNs to near-machine precision, achieving accuracy levels described as equivalent to predicting the diameter of the Earth to within a few centimeters. This level of precision is required for rigorous computer-assisted proofs.

Implications for Fluid Dynamics and Mathematics

Finding singularities in the Navier-Stokes equations is one of the six Millennium Prize Problems. While the current discovery focuses on unstable singularities, it is significant because mathematicians believe no stable singularities exist for boundary-free 3D Euler and Navier-Stokes equations. Consequently, unstable singularities are expected to play a primary role in answering foundational questions about these equations.

As Yongji Wang, first author of the study and Postdoctoral Researcher at NYU, states:

By embedding mathematical insights and achieving extreme precision, we transformed PINNs into a discovery tool that finds elusive singularities.

This work demonstrates a new paradigm for mathematical research where deep mathematical insights are combined with AI to facilitate computer-assisted proofs for long-standing challenges in physics and engineering.

Sources