Consistent and symmetry preserving data-driven interface reconstruction for the level-set method

Aaron B. Buhendwa, Deniz A. Bezgin, Nikolaus A. AdamsJournal of Computational Physics 2022 · DOI · arXiv

In Computational Fluid Dynamics (CFD), accurately tracking the interface between two immiscible fluids—such as air and water in a breaking wave—is a fundamental challenge. Traditional methods rely on geometric reconstructions that often struggle on coarse grids, leading to numerical errors that can compromise the entire simulation. Researchers at the Technical University of Munich have proposed a novel hybrid approach that combines deep learning with classical numerical methods to achieve high accuracy without sacrificing physical consistency.

The Challenge of Interface Reconstruction

The level-set method tracks interfaces using a signed distance field, $\phi$. From this field, solvers must calculate the volume fraction ($\alpha$) and face apertures ($A$) for each computational cell. These values are critical for mass conservation and calculating surface tension forces.

Interface reconstruction parameters in a computational cell, showing level-set values, volume fraction, and face apertures.

Standard linear reconstruction is robust but inaccurate for curved surfaces on coarse meshes, often leading to parasitic currents and errors in surface tension calculation. While neural networks (NNs) offer a potential solution for capturing complex curvatures, they often fail two essential physical requirements:

  1. Symmetry Preservation: Physical results should not change if the coordinate system is rotated or mirrored. NNs are notoriously sensitive to input orientation.
  2. Convergence: As the grid is refined, the error should decrease predictably. Pure NN models often reach an "error floor" where accuracy stops improving, regardless of mesh resolution.

A Hybrid Machine Learning Approach

To address these limitations, the authors developed a "Combined Model" architecture integrated into the ALPACA CFD solver. This system uses a multi-stage process to ensure both accuracy and reliability:

  • The Classifier: A small Convolutional Neural Network (CNN) analyzes the local level-set stencil (e.g., $3 \times 3$) to determine if the interface is simple (linear) or complex (curved). If the interface is simple, the system defaults to classical linear math.
  • The Regression Network: For complex interfaces, a Multilayer Perceptron (MLP) predicts the volume fractions and apertures. The models were trained on synthetic geometric primitives like circles and ellipses but were tested on "star-shaped" interfaces to evaluate generalization.
  • Symmetry Preservation: The model employs test-time augmentation. The input grid is transformed into eight symmetric variations (rotations and reflections). The NN processes all eight variations, and the results are averaged to ensure the final output is perfectly symmetric.
The symmetry-preserving framework utilizes eight grid transformations to ensure physical consistency across rotations and reflections.

Key Insights: Consistency and Convergence

A major breakthrough of this work is the recovery of first-order convergence. By using the classifier to switch back to classical linear reconstruction at high resolutions, the model avoids the "error floor" typical of pure ML approaches. This ensures that the simulation remains mathematically consistent as the mesh is refined, a property essential for scientific computing.

Furthermore, the study found that for the small stencils used in interface reconstruction, simple MLPs are highly effective, often matching or exceeding the performance of more complex CNN architectures when combined with the symmetry-averaging strategy.

Results and Performance

The hybrid model was tested on complex dynamic scenarios, including the "rising bubble" test case. Key findings include:

  • Accuracy: On coarse grids, the NN-based reconstruction reduced the Mean Absolute Error (MAE) by up to an order of magnitude compared to standard linear methods.
  • Mass Conservation: The model demonstrated significantly improved mass preservation in dynamic fluid simulations, addressing a common weakness in level-set methods.
  • Efficiency: Despite the overhead of running the NN multiple times for symmetry, the total simulation time only increased by 11% to 38%. In the context of high-fidelity CFD, this is a manageable trade-off for the substantial gains in accuracy.

Conclusion and Future Directions

This research demonstrates a pragmatic path for integrating machine learning into high-performance CFD. By combining the flexibility of NNs with the rigor of classical methods, the authors have created a tool that is both accurate on coarse grids and physically consistent. Future work will focus on extending these methods to 3D simulations and handling complex topological changes like droplet coalescence and breakup.

Summary updated

BibTeX
@article{buhendwa2022consistent,
  author = {Aaron B. Buhendwa and Deniz A. Bezgin and Nikolaus A. Adams},
  title = {{Consistent and symmetry preserving data-driven interface reconstruction for the level-set method}},
  journal = {Journal of Computational Physics},
  year = {2022},
  doi = {10.1016/j.jcp.2022.111049},
}
Figure