Abstract Large-Eddy Simulation (LES) relies on closure models to represent the influence of unresolved sub-filter scales on resolved flow. While Artificial Neural Networks (ANNs) have emerged as powerful candidates for these closures, their lack of interpretability remains a significant hurdle. This study investigates an ANN-based model for residual scalar flux in homogeneous isotropic turbulence. By employing differential sensitivity analysis, we demonstrate that the ANN implicitly learns a hybrid "mixed model" strategy, combining the physical accuracy of structural models with the numerical stability of functional models. The ANN outperforms classical benchmarks in both static a-priori tests and dynamic a-posteriori simulations, accurately capturing complex phenomena such as back-scatter across multiple filter scales.
Introduction
In fluid dynamics, LES provides a computationally feasible path to simulating turbulence by resolving only large-scale motions. The effect of smaller, sub-filter scales is accounted for via a closure model. Traditional models generally fall into two categories: functional models (e.g., Prandtl Smagorinsky), which provide necessary dissipation but often lack physical fidelity, and structural models (e.g., Scalar Asymptotic Gradient), which are physically grounded but frequently numerically unstable. This work bridges the gap by using deep learning to develop a model that is both accurate and stable, while providing a framework for interpreting its learned physical strategies.
Methodology
The study utilizes high-fidelity Direct Numerical Simulation (DNS) data of a passive scalar in homogeneous isotropic turbulence ($Re_\lambda = 170$, $Pr = 1$). A deep feed-forward ANN with 8 layers and approximately 65,000 parameters was trained to predict the residual scalar flux based on local velocity and scalar field gradients. The model was benchmarked against the Prandtl Smagorinsky (PRS) and Scalar Asymptotic Gradient (SAG) models across four different filter length scales ($\Delta^*$).
Training Dynamics and Statistical Fidelity
A key observation during the training process was the hierarchical nature of learning. The ANN rapidly captures the mean behavior of the target flux but requires significantly more iterations to accurately represent higher-order statistical moments such as variance, skewness, and kurtosis.
As shown in Figure 3, while the training and testing loss (MSE) stabilize relatively early, the higher-order moments (panel b) exhibit more complex convergence behavior, highlighting the necessity of prolonged training to capture the full PDF of the sub-filter flux.
Comparative Performance and Interpretability
In a-posteriori tests, where the model is integrated into a dynamic simulation, the ANN demonstrated remarkable stability and accuracy. Unlike the SAG model, which often diverges, the ANN remained stable over several integral timescales.
Figure 4 illustrates that the ANN consistently achieves lower normalized MSE than both the PRS and SAG models across all investigated filter scales. Sensitivity analysis revealed that the ANN adapts its internal logic based on the resolution: at fine scales, it emphasizes structural reconstruction, while at coarser scales, it increases its functional diffusive component to maintain stability. This "mixed model" behavior allows the ANN to correctly predict back-scatter—the transfer of energy from small to large scales—which is a known limitation of purely functional models like PRS.
Conclusion
This research demonstrates that ANNs are not merely "black boxes" but can be rigorously interpreted to reveal deep physical insights. By naturally discovering a hybrid modeling strategy, the ANN-based closure provides a superior alternative to classical models. Future work will focus on incorporating physical symmetries, such as rotational invariance, directly into the network architecture to further enhance generalizability and robustness in more complex flow regimes.
Summary updated