Abstract
This study presents a novel hybrid approach that combines Graph NeuralNetworks (GNNs) with Reynolds-Averaged Navier Stokes (RANS) equations toenhance the accuracy of mean flow reconstruction across a range of fluiddynamics applications. Traditional purely data-driven Neural Networks (NNs)models, often struggle maintaining physical consistency. Moreover, theytypically require large datasets to achieve reliable performances. The GNNframework, which naturally handles unstructured data such as complex geometriesin Computational Fluid Dynamics (CFD), is here integrated with RANS equationsas a physical baseline model. The methodology leverages the adjoint method,enabling the use of RANS-derived gradients as optimization terms in the GNNtraining process. This ensures that the learned model adheres to the governingphysics, maintaining physical consistency while improving the predictionaccuracy. We test our approach on multiple CFD scenarios, including casesinvolving generalization with respect to the Reynolds number, sparsemeasurements, denoising and inpainting of missing portions of the mean flow.The results demonstrate significant improvements in the accuracy of thereconstructed mean flow compared to purely data-driven models, using limitedamounts of data in the training dataset. The key strengths of this study arethe integration of physical laws into the training process of the GNN, and theability to achieve high-accuracy predictions with a limited amount of data,making this approach particularly valuable for applications in fluid dynamicswhere data is often scarce.