When simulating large-scale fluid systems—such as the Earth's atmosphere, ocean currents, or plasma loops in magnetohydrodynamics—we are limited by grid resolution. We cannot afford to simulate every individual cloud, wave, or small-scale eddy. Instead, we simulate the large, resolved scales directly and use subgrid-scale (SGS) models, or closures, to approximate the collective influence of the unresolved, microscopic dynamics.
For decades, the standard approach has been to treat these unresolved scales as a purely dissipative sponge. Because turbulence naturally breaks large structures down into smaller ones (a forward energy cascade), traditional models use "eddy viscosity" to drain kinetic energy from the resolved scales. While this keeps simulations numerically stable, it introduces a severe systemic flaw: it over-damps the fluid dynamics.
In the real world, turbulence is a bidirectional street. Energy does not just flow downward; it also flows upscale from small-scale fluctuations back into large-scale structures. This phenomenon is known as energy backscatter. Ignoring it causes simulations to lose vital kinetic energy, trapping them in artificial, over-smoothed equilibrium states and failing to capture critical, persistent geophysical structures.
The Cost of Over-Damping
When a model relies solely on eddy viscosity for numerical stability, it systematically underrepresents the system's natural variability. As argued by Chapron et al. in Large‐scale flows under location uncertainty: a consistent stochastic framework, purely dissipative subgrid models over-damp fluid dynamics, trapping simulations near stable equilibrium states and preventing them from exploring the full range of extreme states in the system's natural attractor.
This dissipative bias has severe consequences for long-term forecasting. In climate modeling, for instance, Fogt et al. explain in Stochastic climate theory and modeling that artificial damping degrades the representation of persistent atmospheric structures like blocking events—stationary high-pressure systems crucial for predicting extreme weather. Similarly, in ocean modeling, conventional parameterizations lack the necessary kinetic energy to accurately sustain real-world ocean currents, as noted by Perezhogin et al. in A stable implementation of a data-driven scale-aware mesoscale parameterization.
To fix this, models must allow energy to move bidirectionally between scales. However, doing so deterministically is notoriously difficult; simply adding a negative viscosity term to inject energy typically causes numerical simulations to explode.
Stabilizing Backscatter via Stochastic Forcing
How do we inject energy back into a simulation without causing catastrophic numerical instability? The answer lies in stochastic parameterization—introducing mathematical randomness to represent the chaotic, non-deterministic nature of subgrid interactions.
Instead of trying to predict the exact behavior of every small eddy, stochastic models inject energy as fluctuating noise. As demonstrated by Nadiga et al. in Stochastic parameterization for large eddy simulation of geophysical flows, stochastic forcing allows simulations to maintain the chaotic variability of the flow without the excessive damping associated with traditional approaches. Because the energy injection fluctuates, it averages out safely over time while keeping the large-scale variance alive.
Traditional Closure: [Resolved Scales] ───(Eddy Viscosity)───> [Energy Dissipated]
Stochastic Closure: [Resolved Scales] <───(Random Forcing)───> [Subgrid Fluctuations]
Several mathematical frameworks have been developed to formalize this bidirectional transfer:
- Relaxing Differentiability: Mémin et al. show in STOCHASTIC FLUID DYNAMIC MODEL AND DIMENSIONAL REDUCTION that relaxing differentiability assumptions on the large-scale drift component allows reduced-order fluid models to accommodate bidirectional energy flows rather than assuming a purely dissipative subgrid tensor.
- Memory Effects: By leveraging the Mori-Zwanzig formalism, the authors of A dynamic subgrid scale model for Large Eddy Simulations based on the Mori-Zwanzig formalism derive closures that capture non-local memory effects, successfully predicting energy backscatter in homogeneous turbulence.
- Multiscale Residuals: In magnetohydrodynamics, Sondak et al. demonstrate in A new class of finite element variational multiscale turbulence models for incompressible magnetohydrodynamics that making the turbulence model proportional to the numerical residual permits energy backscatter, which is vital for capturing self-sustaining magnetic fields like the geodynamo.
Machine Learning and the Discovery of Bidirectional Closures
With the rise of high-fidelity simulation data, researchers have increasingly turned to machine learning to capture these complex inter-scale interactions. Neural networks are highly expressive and capable of learning the intricate, non-linear mappings of bidirectional energy transfer directly from data.
In Interpreting neural network models of residual scalar flux, the authors demonstrate that deep learning models naturally "discover" hybrid physics-based structures that represent energy moving from small-scale eddies back to larger scales, overcoming the inherent limitations of traditional dissipative models. Similarly, List et al. show in Learned turbulence modelling with differentiable fluid solvers: physics-based loss functions and optimisation horizons that neural network-based fluid models successfully represent energy backscatter, distinguishing them from purely dissipative subgrid models.
When combined with stochastic frameworks, machine learning can revolutionize large-scale simulations. For example, Zhang et al. show in Implementation and Evaluation of a Machine Learned Mesoscale Eddy Parameterization into a Numerical Ocean Circulation Model that stochastic machine learning models effectively inject energy from small-scale turbulence back into larger-scale flows, allowing ocean simulations to maintain realistic and highly energetic current structures.
The Frontier: Solving the Stability-Accuracy Dilemma
Despite the promise of data-driven backscatter models, a major hurdle remains: the gap between structural accuracy (how well a model predicts subgrid fluxes in isolation) and functional stability (how well the model runs inside a live simulation).
A machine learning model can achieve near-perfect accuracy during offline training, yet cause the simulation to blow up immediately when deployed online. As explored by Jakhar et al. in Learning Closed‐Form Equations for Subgrid‐Scale Closures From High‐Fidelity Data: Promises and Challenges, a key reason for this instability is the inability to perfectly balance inter-scale energy transfer. Because data-driven models often struggle to simultaneously represent energy backscatter and energy dissipation within a single mathematical framework, they can inadvertently inject unbounded energy into specific frequencies, causing numerical explosion (Perezhogin et al.).
The solution is to move away from pure data matching and toward physics-informed AI. By embedding physical constraints directly into the loss functions and architecture, models are forced to respect the delicate balance of energy conservation.
Data-Only Training: Minimize Flux Error ───────> High Accuracy, High Instability
Physics-Informed Training: Minimize Flux Error ┐
Respect Energy Balance ┴─> High Accuracy, Stable Simulation
For instance, Guan et al. demonstrate in Stable a posteriori LES of 2D turbulence using convolutional neural networks: Backscattering analysis and generalization to higher Re via transfer learning that accurate prediction of backscattering, when paired with transfer learning, is critical for achieving stable and accurate large-eddy simulations without over-dissipating the flow.
Ultimately, the field is converging on models that explicitly incorporate these inter-scale energy dynamics. As shown by the authors in Analytical and AI-Discovered Stable, Accurate, and Generalizable Subgrid-Scale Closure for Geophysical Turbulence, incorporating inter-scale energy transfer alongside data matching ensures that the model correctly represents energy flow across different scales. By enforcing these physical boundaries, researchers can finally discover subgrid-scale closures that capture the real-world vitality of energy backscatter while maintaining absolute numerical stability.
Key Foundational Work
Stochastic parameterization for large eddy simulation of geophysical flows (2006) — contributed 3 concept elements to this direction.
Stochastic climate theory and modeling (2014) — contributed 2 concept elements to this direction.