Recovering missing CFD data for high-order discretizations using deep neural networks and dynamics learning

Kevin T. Carlberg, Antony Jameson, Mykel J. Kochenderfer, Jeremy Morton, Liqian Peng, Freddie D. WitherdenJournal of Computational Physics 2019 · DOI

Introduction

High-fidelity Computational Fluid Dynamics (CFD) simulations are indispensable for the design and optimization of complex systems in aerospace and turbine engineering. Modern high-order discretization methods, such as Discontinuous Galerkin (DG), provide exceptional accuracy but at a significant cost: the "data wall." These simulations generate petabytes of data, far exceeding the practical limits of storage and the bandwidth of I/O systems.

To manage this, engineers typically save only a small fraction of the simulated time steps. However, if a specific intermediate time instance is later identified as critical for analysis, it is effectively lost, necessitating a costly re-run of the entire simulation. This paper proposes a novel two-stage machine learning framework for a posteriori data recovery, allowing researchers to accurately "fill in the blanks" between saved snapshots without the need for additional simulation runs.

Methodology

The authors introduce a hierarchical approach that combines deep learning for spatial compression with regression techniques for temporal dynamics. This strategy is designed to handle the massive dimensionality inherent in industrial-scale physics.

1. Hierarchical Dimensionality Reduction

Standard neural network architectures cannot directly process the global state of a simulation with millions of degrees of freedom due to memory constraints. The authors circumvent this by implementing a two-tier compression strategy:

  • Local Compression (Autoencoders): Deep convolutional autoencoders are applied to each individual element in the mesh. By focusing on the element level, the model leverages the high-order structure of the data and benefits from a massive increase in training samples (total time steps multiplied by the number of elements).
  • Global Compression (PCA): The local encodings from all elements are aggregated and further compressed using Principal Component Analysis (PCA). This creates a compact, low-dimensional "latent state" that represents the entire physical system.

2. Dynamics Learning

The evolution of this latent state is treated as a discrete-time dynamical system. The authors evaluated several regression techniques to learn the transition between states, including:

  • Support Vector Regression (SVR)
  • Random Forests
  • Sparse Identification of Non-linear Dynamics (SINDy)
  • Vectorial Kernel Orthogonal Greedy Algorithm (VKOGA)

Experimental Setup

The framework was evaluated using a large-scale, industrially relevant case: a 3D simulation of flow over an extruded half-cylinder.

  • Parameters: Reynolds number 350 and Mach 0.2.
  • Scale: The mesh consisted of 40,584 elements, resulting in approximately 13 million degrees of freedom.
  • Temporal Scope: 8,600 time instances were used to train and validate the recovery process.

Key Results and Insights

Extreme Compression and Accuracy

The hierarchical framework achieved a staggering 26,000:1 compression ratio, reducing the 13-million-dimensional state to a latent vector of just 500 components. Despite this reduction, the model recovered the global solution with sub-1% relative error across the entire temporal sequence.

Physical and Visual Fidelity

Beyond numerical error metrics, the model demonstrated high physical fidelity. It accurately reconstructed integrated quantities like lift and drag forces, which are notoriously sensitive to minute errors in the flow field. Qualitative visualizations using Q-criterion iso-surfaces confirmed that the reconstructed flow captured all essential vortical structures present in the original high-fidelity data.

The Importance of Stability

A critical insight from the study involves the stability of the dynamics learning phase. While some methods like SINDy showed low error on static test snapshots, they often produced unstable results when integrated over time because their underlying functions were unbounded. The researchers found that methods utilizing Gaussian kernels (such as VKOGA and SVR with RBF kernels) were inherently bounded, providing much more stable and reliable long-term reconstructions.

Conclusion and Future Work

This research demonstrates that a "local-to-global" machine learning strategy can effectively overcome the data storage challenges of high-order CFD. By making deep learning tractable for massive-scale physics, the authors provide a tool for efficient data recovery and analysis.

Future research will focus on improving the generalization of these models across different physical parameters (such as varying Reynolds or Mach numbers) and developing mesh-agnostic architectures to remove the dependency on specific simulation grids.

Summary updated

BibTeX
@article{carlberg2019recovering,
  author = {Kevin T. Carlberg and Antony Jameson and Mykel J. Kochenderfer and Jeremy Morton and Liqian Peng and Freddie D. Witherden},
  title = {{Recovering missing CFD data for high-order discretizations using deep neural networks and dynamics learning}},
  journal = {Journal of Computational Physics},
  year = {2019},
  doi = {10.1016/j.jcp.2019.05.041},
}

Loading related content...

Figure