Scientific Ml

Neural Closure Models

Neural closures learn the unresolved subgrid terms of a coarse simulation, letting a cheap grid approximate expensive high-resolution physics.

The closure problem

When continuous physics is averaged or coarsened, the resulting equations contain terms that depend on scales the coarse model does not resolve. These unknown terms must be expressed through the resolved variables to close the system. Turbulence is the classic example: the effect of small eddies on the large-scale flow appears as a stress that the coarse equations cannot compute directly.

Closures learned from data

Kronos motion — grid 2040

Traditional closures are hand-crafted approximations with limited accuracy. A neural closure instead learns the unresolved term as a function of the resolved state, trained on high-fidelity data or on filtered fully resolved simulations. The learned closure can capture patterns that simple algebraic models miss, improving the coarse simulation without paying for full resolution.

Training against the solver

A closure trained only to match instantaneous subgrid terms often destabilizes long simulations. Better results come from training the closure inside the coarse solver over many time steps, so it is optimized for the trajectories it will actually produce. This solver-in-the-loop training, which requires a differentiable solver, yields closures that stay stable over long rollouts.

Guarding physics

Enforce conservation so the closure cannot create or destroy conserved quantities
Respect symmetries such as rotation and reflection through the architecture
Bound the closure's magnitude to prevent runaway feedback
Fall back to a physical closure where inputs leave the training regime

Outlook

Neural closures are among the most promising uses of scientific machine learning because they attack a genuine bottleneck: the cost of resolving small scales. They are surrogates for expensive physics, validated against resolved simulations, and are used to extend the reach of coarse models rather than to assert new physical results.