Consistency requirement of data-driven subgrid-scale modeling in large-eddy simulation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Consistency requirement of data-driven subgrid-scale modeling in large-eddy simulation".
Title and authors: Kai: So to recap what we just covered about this paper on the "Consistency requirement of data-driven subgrid-scale modeling in large-eddy simulation," we're talking about how numerical errors mess up the training process for these AI models.
Mira: Exactly. The authors are examining this inconsistency between a priori and a posteriori tests by looking at two different model types, the eddy viscosity and complex nonlinear models, to see which one handles the numerical deviation better.
Lev: It sounds like they're trying to figure out what's causing that mismatch, because if you can fix that underlying numerical difference, you stop the training from becoming unreliable.
Kai: They found that by incorporating those specific numerical deviations into the training data itself, it helps stabilize simulations and gets better predictions for energy spectra when they test the model later.
Mira: That brings us to how they build this data—they use a DNS-aided LES framework where they run filtered direct numerical simulation and LES side-by-side to directly measure these discrepancies.
Lev: So, the core idea is that if you train on data that already has those known numerical errors built in, the resulting model should be more robust when you try to apply it to a new situation.
The paper's summary: Kai: To summarize what this paper on "Consistency requirement of data-driven subgrid-scale modeling in large-eddy simulation" is saying, they are addressing the difficulty of training accurate SGS models because of that a priori versus a posteriori inconsistency.
Mira: They detail the problem by showing that numerical derivatives of filtered quantities in LES often don't match the ones computed from high-resolution data, and this mismatch comes from commutation errors and numerical errors, which they call delta mom and delta div.
Lev: So they're basically saying that when you try to model turbulence using data-driven methods, you have to account for these specific numerical discrepancies caused by the filter choice and the scheme you use.
Kai: They test two model forms: the eddy viscosity model, which is simpler, and a complex nonlinear model that has more terms, like delta mom one through four.
Mira: The results show that while models including these numerical deviations can sometimes have lower root mean square errors during training, they might also have weaker correlation compared to models that ignore those deviations.
Lev: But the key result is with the complex nonlinear model; it shows significantly improved correlation between its predictions and the true dissipation values.
The paper's improvements: Kai: The paper suggests a few ways to improve this situation, focusing on how we train these models to handle those inconsistencies.
Mira: They highlight that the importance of accounting for numerical deviation really increases in data-driven approaches because they introduce significant nonlinearity and equation coupling, which makes them more sensitive to these errors.
Lev: The paper confirms that excluding the numerical deviation leads to a simulation where kinetic energy grows way beyond what we expect, which means it severely overpredicts small-scale turbulence when the model form allows it.
Kai: Including the numerical deviation instead results in a modest suppression of small-scale energy, which leads to better alignment with the filtered DNS data in those a posteriori tests.
Mira: The conclusion they draw is that this approach of including the numerical deviation makes predictions more physically consistent, and it's crucial for achieving consistency between the a priori and a posteriori tests.
Conclusion: Kai: So to wrap up this paper on "Consistency requirement of data-driven subgrid-scale modeling in large-eddy simulation," the main point is that incorporating numerical deviation into model training is essential for stability and accuracy.
Mira: They showed that excluding it leads to exaggerated small-scale turbulence predictions, while including it yields more physically consistent results across both testing phases.
Lev: For someone building real hardware, this means you need to be careful; if the model isn't trained with these corrections, the solution can drift away from the valid state space when nonlinear effects take over.
Kai: The paper confirms that this method of including numerical deviation results in more physically consistent predictions for both energy spectra and kinetic energy levels.
Mira: It seems like this approach is crucial for achieving consistency between a priori and a posteriori tests in data-driven SGS modeling, which is what the authors were really aiming to show.
Lev: So we're seeing that these AI models can be made more reliable when they are trained on data that acknowledges the numerical reality of the simulation setup.
Graduate Aerospace Laboratories, California Institute of Technology · Department of Earth, Environmental & Planetary Sciences, Brown University · Department of Mechanical and Civil Engineering, California Institute of Technology
physics.flu-dyn
Submitted: 2025-11-19
Updated: 2025-11-19
Comments: 16 pages, 10 figures
DOI: 10.1103/yykb-6rvf
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
The gist: Data-driven subgrid-scale (SGS) modeling in large-eddy simulations suffers from inconsistency between a priori tests and a posteriori tests…
Key concepts
- Inconsistency Origin
- The mismatch between results from initial tests (a priori) and final tests (a posteriori) stems from errors when numerical differentiation and spatial filtering are combined. Specifically, the way derivatives are calculated numerically often differs significantly from how they appear in high-resolution data, creating a discrepancy that complicates model training.
- Numerical Deviation ($\delta$mom and $\delta$div)
- These represent the specific errors introduced by the numerical schemes used in LES. They quantify the difference between theoretical derivatives and those computed from actual filtered quantities. These deviations depend on the chosen filter and numerical method, making them a key factor that needs to be incorporated into model training for better accuracy.
- Eddy Viscosity Model ($\tau_{ev}$)
- This is one of the two mathematical forms used to describe the subgrid-scale stress. It relates the SGS stress directly to the filtered velocity gradients using a simple relationship. When trained with numerical deviations, it shows lower training errors but sometimes weaker correlation than models without them.
- Complex Nonlinear Model ($\tau_{cn}$)
- This is a more sophisticated model form that incorporates multiple terms involving different orders of the strain-rate tensor (S) and rotation-rate tensor (R). Training this model with numerical deviations resulted in significantly improved correlation between its predictions and the true dissipation values, suggesting it better captures complex turbulence dynamics.
Terminology
Summary
Data-driven subgrid-scale (SGS) modeling in large-eddy simulations suffers from inconsistency between a priori tests and a posteriori tests, which makes training accurate SGS models a difficult task The impact of the numerical deviation is examined through two SGS model formulations: the eddy-viscosity and the complex nonlinear models By incorporating numerical deviations into model training, this approach enhances consistency, stabilizes simulations, and improves predictions of energy spectra in a posteriori tests Our findings highlight that data-driven methods introduce significant nonlinearity and equation coupling, exacerbating inconsistencies compared to non-data-driven approaches. p. 0 By incorporating numerical deviations into model training, we enhance consistency,
stabilize simulations, and improve predictions of energy spectra in a posteriori tests. Our findings highlight
that data-driven methods introduce significant nonlinearity and equation coupling, exacerbating inconsistencies
compared to non-data-driven approaches.Where the paper says it
Inconsistency Origin
The inconsistency between a priori and a posteriori results arises from commutation errors between discretization and differentiation operators, as well as the coarse spatial resolution inherent to LES Numerical derivatives of filtered quantities in LES often differ significantly from the filtered derivatives computed from high-resolution data. This mismatch is due to numerical deviations denoted by δmom and δdiv, which depend on the specific choice of filter and numerical scheme. p. 1 The extra terms δ
mom
i
and δ
div are the numerical deviation that
differentiate the a priori tests from the a posteriori tests. Consequently, the actual values are dependent on the specific
choice of the filter and the numerical scheme.Where the paper says it
Training Signal Construction
The training data is generated using a DNS-aided LES framework where filtered DNS and LES are run side-by-side to directly evaluate discrepancies The dataset used for model development is obtained from the DNS-aided LES described in Section II A. For training, the model form follows the physical constraint of Galilean invariance, involving strain-rate tensor S and rotation-rate tensor R. p. 4 To achieve generalizability, we use the model
form that follows the physical constraint of Galilean invariance. The model involves the strain-rate tensor S and the
rotation-rate tensor R directly computed from the LES flow field, where
Sij =
1
2
∂ui
∂xj
+
∂uj
∂xi
, Rij =
1
2
∂ui
∂xj
−
∂uj
∂xi
. (7)
At the same time, the four tensor invariants are also linear or nonlinear forms of S and R [39],
m1 = S, m2 = S
2
,
m3 = R2
, m4 = SR − RS,
(8)
In the present work, two major model forms are used: the eddy viscosity model
τ ev = −2νT (I1, . , I6)m1, (9)
and the complex nonlinear model
τ cn = − 2νT ,1(I1, .Where the paper says it
Where the paper says it
Impact Analysis
The impact of the numerical deviation is examined in both a priori and a posteriori tests.
** A priori training performance shows that models including the numerical deviation can exhibit lower root mean square errors (rRMSE) but sometimes weaker correlation compared to models excluding it, especially for the eddy-viscosity model The complex nonlinear model trained with the numerical deviation shows significantly improved correlation between predictions and true dissipation values. p. 7 Notably, the correlation between predictions and true values, shown in Fig. 4(f), is significantly
improved compared to the eddy-viscosity model.Where the paper says it
** A posteriori results reveal that excluding the numerical deviation leads to a simulation where kinetic energy grows well beyond the expected level, indicating severe overprediction of small-scale turbulence Including the numerical deviation leads to a modest suppression of small-scale energy, resulting in better spectral alignment with filtered DNS data. 9**
Generalizability and Conclusion
The importance of accounting for numerical deviation is heightened in data-driven approaches due to their inherent nonlinearity The comparison confirms that excluding the numerical deviation leads to exaggerated small-scale turbulence predictions when the model form allows it. Including the numerical deviation results in more physically consistent predictions. p. 0 Our findings highlight
that data-driven methods introduce significant nonlinearity and equation coupling, exacerbating inconsistencies
compared to non-data-driven approaches. Finally, while the impact of the numerical deviation can be generalized, achieving accurate model predictions necessitates a physically grounded model form and an optimal filter
width. Based on
these considerations, we explore two representative classes of SGS models in this study: an eddy-viscosity model and
a complex nonlinear model. In this work, we investigate the consistency between a priori and a posteriori tests in data-driven SGS modeling.Where the paper says it
Where the paper says it
Appendix
The training history of the examined models shows that the convergence of the loss function for both training and validation datasets indicates that the training is finished without overfitting. All four models have shown a converged performance after approximately 100 epochs. p. 5 Training is terminated when the loss function plateaus,
which typically occurs after approximately 100 epochs. The training history of the models is shown in the appendix.Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Inclusion of numerical deviation in model training enhances consistency by incorporating
the numerical deviation into model training, we enhance consistency, stabilize simulations, and improve predictions of energy spectra in a posteriori tests.
This allows data-driven methods to better align with the governing equations encountered during LES deployment. -
The AI system can perform more physically realistic a posteriori simulations by using models trained with the numerical deviation instead of those excluding it. This leads to
a priori and a posteriori kinetic energy spectra are generally comparable or even superior,
which is critical for stable LES performance. -
The improved system can better handle complex flow scenarios by utilizing richer model forms, such as the complex nonlinear model, when trained with numerical deviation. This helps the model
better reconcile multiple scales of turbulence
compared to models lacking this correction. -
The AI system will exhibit improved stability in complex flows because incorporating numerical deviation
eases the challenge of such inconsistency between a priori and a posteriori tests in data-driven modeling.
This prevents solutions from drifting away from the valid state space when nonlinear effects dominate. -
The system can adapt to different filter widths and grid resolutions while maintaining consistent performance by using an intermediate scale for the filtered DNS filter width. This ensures
satisfactory a posteriori performance
across varying numerical setups, which is a key trade-off identified in the study.
Abstract
Data-driven subgrid-scale (SGS) modeling in the large-eddy simulations (LES) suffers from the inconsistency between the a priori tests and the a posteriori tests, which make training accurate SGS models a difficult task. We study the difference in filtered high-fidelity data and LES to identify the numerical deviation between the two cases, which is a combined impact of commutation error, numerical errors, and error coupling. The impact of the numerical deviation is examined through two SGS model formulations: the eddy-viscosity and the complex nonlinear models. By incorporating numerical deviations into model training, we enhance consistency, stabilize simulations, and improve predictions of energy spectra in a posteriori tests. Our findings highlight that data-driven methods introduce significant nonlinearity and equation coupling, exacerbating inconsistencies compared to non-data-driven approaches. Finally, while the impact of the numerical deviation can be generalized, achieving accurate model predictions necessitates a physically grounded model form and an optimal filter width.
Sources
- Numerical and modeling error assessment of large-eddy simulation using direct-numerical-simulation-aided large-eddy simulation
- Adam: A Method for Stochastic Optimization
Related papers
- Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence
- Early onset of secondary shear instability in Kelvin-Helmholtz braids at high Reynolds number
- Rotational covariance, continuous spectrum and non-modal growth of three-dimensional planar weak viscous shock layers
- Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows
- Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows
- Size and shape of condensing steam jets and plumes