MAgNET for CFD: Evaluation and Application

Type of thesis: Masterarbeit / location: Dresden / Status of thesis: Open theses


In the aerodynamic design process, many hundreds to thousands of computationally intensive
Computational Fluid Dynamics (CFD) simulations are performed to find an optimal design for the given boundary conditions. To reduce the time and resource requirements, surrogate models are typically used, allowing a significant portion of CFD calculations to be saved. The meaningfulness and flexibility of these surrogate models have significantly increased in recent years through the use of machine learning (ML) methods.
Graph-based ML algorithms, which avoid lossy transformations of the numerical grid, have gained
importance in recent years [1-4] and have shown that predicting flow parameters directly on the
numerical grid is possible. In this thesis, a model architecture’s applicability to CFD data, which has
previously only been applied to structural mechanical problems, will be investigated. The goal is to
evaluate the model framework MAgNET [5,6] and apply it to fluid dynamic data for predicting flow
fields for unknown geometries.
After a detailed description of the model architecture, at least one successful training on at least one two-dimensional fluid dynamics dataset [7] should be carried out. Furthermore, differences and similarities between MAgNET and comparable models [1,4] should be highlighted, and the advantages and disadvantages of each architecture should be discussed. Finally, one of the ML frameworks [1,4] should be selected, and a comparable training should be set up. The prediction accuracy, speed, and resource consumption of both models should then be compared using suitable metrics.

Thematic Focus

  • Literature Research
  • Detailed breakdown of the ML framework MAgNET
  • Setting up a training pipeline for training MAgNET on CFD data
  • Setting up a training pipeline for training [1] or [4] on CFD data
  • Comparison of the training and validation of both models using appropriate metrics
  • Additional: Analysis of training and prediction performance with Score-P + Vampir


[1] M. Fortunato, T. Pfaff, P. Wirnsberger, A. Pritzel, and P. Battaglia, “MultiScale MeshGraphNets,” presented at
the ICML 2022 2nd AI for Science Workshop, Jul. 2022.
[2] M. Lino, A. A. Bharath, S. Fotiadis, and C. Cantwell, “Towards Fast Simulation of Environmental Fluid
Mechanics with Multi-Scale Graph Neural Networks,” ICLR22, 2022.
[3] S. Strönisch, M. Sander, M. Meyer, and A. Knüpfer, “Multi-GPU Approach for Training of Graph ML Models on
large CFD Meshes,” in AIAA SCITECH 2023 Forum, AIAA, Jan. 2023. doi: 10.2514/6.2023-1203.
[4] Y. Cao, M. Chai, M. Li, and C. Jiang, “Efficient Learning of Mesh-Based Physical Simulation with Bi-Stride
Multi-Scale Graph Neural Network,” ICML23, Jun. 2023
[5] Deshpande, S. (2023). „Data Driven Surrogate Frameworks for Computational Mechanics: Bayesian and
Geometric Deep Learning Approaches“. Doctoral thesis, Unilu – University of Luxembourg


Sebastian Strönisch