Engineering · 8 min read · April 28, 2026
Learning turbulence closures via nudging sidesteps solver backprop
A data-assimilation-inspired approach trains neural network turbulence models on DNS data without embedding them in solvers, reducing computational cost and improving stability.
Source: arxiv/cs.LG · Ashwin Suriyanarayanan, Melissa Adrian, Dibyajyoti Chakraborty, Romit Maulik · open original ↗ ↗
Nudging-based training lets neural turbulence closures learn from DNS data without costly solver backpropagation or stability issues.
- — A-posteriori learning embeds neural closures in solvers but requires expensive gradient backpropagation and causes instability.
- — A-priori learning uses DNS data directly but assumes filter properties that don't match actual numerical discretization effects.
- — Continuous data assimilation (nudging) treats DNS as sparse observations and trains closures offline without modifying the solver.
- — Nudging approach avoids adjoints, reduces computational burden, and maintains long-term stability in LES deployments.
- — Model generalizes across different numerical schemes and temporal discretizations better than traditional closure models.
- — No need to embed neural network inside solver, lowering barrier to adoption in existing simulation codes.
- — Addresses mismatch between assumed filter properties and real numerical discretization errors that destabilize standard approaches.
Frequently asked
- Nudging, or continuous data assimilation, is a technique that treats high-fidelity DNS data as sparse observations and uses a forcing term to guide a coarse-grid model toward those observations. In this work, it allows a neural network closure to learn the required subgrid stress without being embedded inside the LES solver, reducing computational cost and avoiding stability issues from filter mismatch.