Learning Closed-form Equations for Subgrid-scale Closures from High-fidelity Data: Promises and Challenges

Overview

• Researchers are exploring ways to discover interpretable, closed-form equations for subgrid-scale (SGS) processes in complex systems like the Earth's climate.
• This paper applies an equation-discovery technique to learn SGS closures from simulations of 2D turbulence and Rayleigh-Bénard convection (RBC).
• The discovered closures are consistent with the nonlinear gradient model (NGM), which can be derived analytically.
• While the discovered closures show strong correlations with the true fluxes, large-eddy simulations using NGM closures are unstable, due to limitations in the model.
• The paper proposes ideas for learning more accurate and stable SGS closures in future work.

Plain English Explanation

When scientists study complex systems like the Earth's climate, they often use simplified mathematical models to represent the small-scale processes that can't be directly captured. These simplified models are called subgrid-scale (SGS) closures or parameterizations.

In this research, the scientists used a technique called "equation discovery" to try to find simple, interpretable mathematical equations that could describe these SGS processes. They applied this technique to computer simulations of two common fluid flow problems: 2D turbulence and Rayleigh-Bénard convection (where hot fluid rises and cool fluid sinks, creating circulation patterns).

Interestingly, the equation-discovery process consistently found the same type of mathematical relationship, known as the nonlinear gradient model (NGM). This model describes the SGS fluxes (the transfer of momentum or heat between large and small scales) in terms of the gradients, or changes, of the large-scale variables.

The scientists found that the NGM closures had a strong correlation with the true SGS fluxes from the simulations. However, when they used these NGM closures in large-scale simulations, the simulations became unstable and crashed.

The researchers identified two key reasons for this instability. First, in 2D turbulence, the NGM model did not properly capture the transfer of energy between the large and small scales. Second, in the Rayleigh-Bénard convection, the NGM model did not accurately predict the backflow of heat from the small to large scales.

To address these issues and develop more accurate and stable SGS closures, the researchers propose several ideas. These include using physics-informed libraries of potential models, modifying the loss functions and metrics used during the discovery process, and accounting for the uncertainty and non-uniqueness of the SGS fluxes used as the "ground truth" data.

Overall, this research highlights the challenges in discovering interpretable mathematical models for the small-scale processes in complex systems. While the equation-discovery technique was able to find a common form for the SGS closures, there are still significant hurdles to overcome in order to develop models that are both accurate and numerically stable.

Technical Explanation

The researchers applied an equation-discovery technique, using expansive libraries of potential mathematical expressions, to learn SGS closures from filtered direct numerical simulations (DNS) of 2D turbulence and Rayleigh-Bénard convection (RBC).

Across common filtering approaches (e.g., Gaussian, box), the technique robustly discovered closures for the momentum and heat fluxes that took the same general form. These closures depended on nonlinear combinations of gradients of the filtered variables, with constants that were independent of the fluid/flow properties and only depended on the filter type and size.

The discovered closures were found to be consistent with the nonlinear gradient model (NGM), which can be derived analytically using Taylor series expansions. This suggests that for many common systems and physical processes, equation-discovery algorithms tend to converge on the leading-order Taylor series term as the learned closure model, except when using certain types of filters (e.g., cutoff filters).

However, despite the strong correlations (>0.95) between the true and NGM-predicted fluxes, the researchers found that large-eddy simulations using the NGM closures were unstable. They identified two key shortcomings of the NGM model:

1. In 2D turbulence, the NGM closure produces zero kinetic energy transfer between the resolved and subgrid scales, lacking both diffusion and backscattering effects.
2. In RBC, the NGM closure poorly predicts the potential energy backscattering from small to large scales.

Furthermore, the researchers showed that the SGS fluxes diagnosed from the simulation data, which were used as the "ground truth" for the equation discovery, depend on the filtering procedures and are not unique.

To address these issues and develop more accurate and stable SGS closures in future work, the researchers propose several ideas. These include using physics-informed libraries of potential models, modifying the loss functions and metrics used during the discovery process, and accounting for the uncertainty and non-uniqueness of the SGS fluxes.

Critical Analysis

The researchers provide a thorough and thoughtful analysis of the limitations of the equation-discovery approach for learning SGS closures, as demonstrated by the instability of the NGM closures in large-eddy simulations.

They correctly identify the key shortcomings of the NGM model, namely its inability to capture important subgrid-scale phenomena like energy transfer and backscattering. This is a significant insight, as it highlights the need for more sophisticated closure models that can account for these complex physical processes.

The researchers also raise an important point about the non-uniqueness of the SGS fluxes used as the "ground truth" data for the equation-discovery process. This is a critical issue, as the accuracy and stability of the learned closures will inherently depend on the quality and robustness of the training data.

While the researchers propose several promising ideas for addressing these challenges, such as using physics-informed libraries and modified loss functions, it remains to be seen how effective these approaches will be in practice. Further research and validation will be necessary to determine the best path forward for developing accurate and stable SGS closures using equation-discovery techniques.

Conclusion

This research highlights the ongoing challenge of discovering interpretable, closed-form equations for subgrid-scale processes in complex systems like the Earth's climate. While the equation-discovery technique was able to consistently identify the nonlinear gradient model as a common closure form, the researchers found that this model suffers from significant limitations in terms of numerical stability and the ability to capture key physical phenomena.

To address these issues, the researchers propose several ideas for future work, including the use of physics-informed libraries, modified loss functions and metrics, and a more rigorous treatment of the uncertainties and non-uniqueness inherent in the SGS flux data. Implementing and validating these approaches could lead to more accurate and stable SGS closures, which would have important implications for a wide range of fields that rely on coarse-grained modeling of complex systems.