Mapping CFD Simulations to ML Convolutions

In the age of Machine Learning, data-driven reduced-order models (ROMs) are gaining attraction across numerous scientific disciplines. Convolutional Autoencoders, in particular, met considerable success for both data compression and surrogate modeling [1]. When handling scientific datasets, model architectures are often tailored to the data’s underlying structure (e.g. graph neural networks for unstructured computational grids, and convolutional networks for structured data). While standard convolutional layers offer computational efficiency and straightforward training, their applicability is inherently limited to data organized on regular grids [2]. However, Computational Fluid Dynamics (CFD) simulations of complex geometries, are typically performed on unstructured meshes, which cannot be directly processed by such layers. In this presentation, we introduce three methods to deal with unstructured CFD surface data: (i) current most adopted procedure of direct interpolation of unstructured data onto a predefined structured grid, (ii) the use of a conformal mapping technique [3] to transform unstructured CFD meshes into structured computational domains and (iii) a grid mapping based on the shortest distance path of the computational graph built on the CFD grid. Pros and cons of each algorithm will be discussed. The method will be applied to CFD solutions on swept wings, varying the sweep angle and the angle of attack.