A novel compact scheme for second-order fluxes applied to the Spectral Difference method

This paper presents a novel compact scheme for the discretization of second-order (viscous) fluxes within the Spectral Difference (SD) method, a type of Discontinuous Spectral Element Method (DSEM). The work is motivated by the deficiencies of the standard Bassi-Rebay (BR1) formulation, which, while simple and parameter-free, suffers from sub-optimal convergence at even polynomial orders and requires an extended five-element computational stencil in one dimension.

The proposed approach, inspired by Huynh's work on Flux Reconstruction, modifies the reconstruction of the auxiliary gradient variable by introducing interface-dependent, one-sided continuous fluxes. This reduces the stencil for the viscous discretization from five to three elements while preserving the centered and parameter-free nature of the original BR1 formulation. The methodology is detailed for the one-dimensional case and extended to multiple dimensions using the tensor-product structure of SD elements.

A temporal eigenanalysis, including a combined-mode formulation and the addition of interior penalty (IP) terms, is used to characterize the dissipation and dispersion properties. The analysis shows that the compact formulation is generally more dissipative at high wavenumbers, particularly at low polynomial orders, and that differences between the two schemes diminish as the IP parameter increases.

Numerical experiments are conducted, including:

1. Convergence study for linear diffusion: The compact formulation recovers the expected convergence order for all polynomial degrees (p=1 to 5), while the standard BR1 scheme exhibits degraded, sub-optimal convergence at even orders (p=1, 3, 5). For example, at p=1, the compact scheme achieves a rate of 2.599 versus 0.998 for the standard, and at p=3, 4.186 versus 2.996.

2. Diffusion of an under-resolved Dirac's delta: The compact scheme significantly reduces spurious oscillations and periodically excited high-wavenumber modes compared to the standard formulation, consistent with the eigenanalysis predictions.

3. Nonlinear porous medium equation (PME): The compact scheme consistently yields smaller L2 errors and reduced oscillatory behavior. The addition of IP terms is beneficial mainly for the standard formulation and can even be detrimental to the compact one at higher resolutions.

4. Implicit Large-Eddy Simulations (ILES) of the 3D Taylor-Green vortex (TGV) at Re = 1,600 and 5,000: The compact scheme remains stable in several configurations where the standard BR1 formulation fails, due to its improved damping of spurious high-wavenumber modes. In stable cases, the compact scheme produces lower levels of small-scale enstrophy activity.

The paper concludes that the proposed compact formulation is an effective and computationally attractive alternative to the standard BR1 scheme, particularly for large-scale and implicit LES applications, due to its reduced stencil, improved accuracy, and enhanced robustness in under-resolved flows.

Improvements to AI Systems:

1. Enhanced Neural PDE Solvers with Compact Stencil Regularization:

• Improvement: Integrate the compact one-sided flux reconstruction (inspired by Huynh) into neural operators (e.g., Fourier Neural Operators, U-Nets) used for solving viscous PDEs. Replace the standard BR1-like gradient estimation in the neural network’s discretization layer with the proposed three-element stencil.

• What the improved AI can do: Achieve higher-order convergence for even polynomial degrees in learned solvers, reduce spurious oscillations in under-resolved turbulence predictions, and maintain stability in long-horizon rollouts without requiring extra dissipation or artificial viscosity.

2. Stability-Aware Training for Implicit LES Models:

• Improvement: Use the eigenanalysis results (dissipation/dispersion spectra) to design a loss function that penalizes the neural surrogate for matching the unstable high-wavenumber modes of BR1. Train the AI model to emulate the compact scheme’s damping characteristics.

• What the improved AI can do: Produce stable and physically accurate implicit large-eddy simulations (e.g., Taylor-Green vortex at Re=5,000) even when the training data is limited or under-resolved, avoiding blow-up and reducing small-scale enstrophy errors.

3. Adaptive Flux Selection in Learned Discretizations:

• Improvement: Train a meta-learning AI to dynamically choose between BR1 and the compact scheme based on local flow features (e.g., wavenumber content, polynomial order, IP parameter). Use the paper’s convergence and stability data as training labels.

• What the improved AI can do: Automatically switch to the compact scheme in high-gradient regions to suppress oscillations, and revert to BR1 in smooth regions to save computation, optimizing both accuracy and cost in real-time simulations.

4. Reduced-Stencil Graph Neural Networks for Mesh-Based Solvers:

• Improvement: Modify graph neural network message-passing to use the three-element stencil (instead of five) for viscous flux computation, as proposed in the paper. This reduces the receptive field and communication overhead.

• What the improved AI can do: Scale to larger 3D meshes with lower memory and latency, while preserving high-order accuracy—critical for GPU-accelerated, real-time digital twin applications.

5. IP-Parameter-Aware Uncertainty Quantification:

• Improvement: Train an AI to predict the optimal interior penalty (IP) parameter for the compact scheme based on resolution and polynomial order, using the paper’s finding that IP is less beneficial (even detrimental) for the compact scheme at high resolution.

• What the improved AI can do: Automatically tune numerical schemes in a solver-agnostic manner, avoiding over-dissipation and ensuring robust performance across different mesh densities and flow regimes.

6. Generative Data Augmentation for Under-Resolved Flow Prediction:

• Improvement: Use the compact scheme’s superior damping of Dirac-delta diffusion (experiment 2) to generate clean, high-fidelity training data for AI-based super-resolution or turbulence closure models.

• What the improved AI can do: Learn to reconstruct fine-scale features from coarse inputs without introducing artificial noise, improving the accuracy of downstream predictions (e.g., drag coefficients, mixing rates) in industrial CFD.