Reproduce “Machine Learning-Based Optimal Mesh Generation in Computational Fluid Dynamics” on OpenResearch

About this paper

Keefe Huang, Moritz Krügener, Alistair Brown, Friedrich Menhorn, Hans-Joachim Bungartz, Dirk Hartmann

Researchers from Technical University of Munich and Siemens developed a machine learning framework that predicts optimal mesh densities for computational fluid dynamics (CFD) simulations. This framework leverages a large dataset of adjoint-based adaptive mesh refinement outputs to train a convolutional neural network, achieving 98.7% accuracy in predicting mesh distributions for efficient simulations.

Key insights

• Machine learning can effectively encapsulate the knowledge of computationally demanding, adjoint-based AMR techniques into a predictive model, making advanced mesh optimization accessible without direct adjoint computations.
• Predicting mesh density as a grayscale image on a rectilinear grid allows a convolutional neural network to learn complex, spatially varying refinement patterns for diverse geometries.
• Applying machine learning to mesh generation, rather than directly replacing flow solvers, reduces industrial validation concerns as a sub-optimal mesh primarily affects computational cost, not the physical correctness of the simulation results.

Summary by alphaXiv.