Abstract The development of 3-dimensional airfoils for turbomachines using Computational Fluid Dynamics (CFD) requires significant time and resources. This is due to the iterative nature of the design process, where new meshes and CFD simulations must be run for each design iteration. The integration of Artificial Intelligence (AI) offers a solution by enabling the creation of surrogate models for the flow field based on a training dataset. We investigate various methods to build these surrogates, in the context of axial compressor design. The baseline approach predicts the flow field as an average of the n most similar cases in the training set. The field compression computes latent representations of the geometry and field, and the predictions are made in the latent space. Finally, we test MeshGraphNet, a GNN (Graph Neural Network)-based approach from DeepMind. We compare the performance of these models in predicting the flow field in a single rotor, as well as the KPI (Key Performance Indicators) for compressor design. Furthermore, we evaluate the trade-off between training dataset size and model performance to optimize the business value of these AI methods. This assessment underscores the importance of minimizing computational and human time costs while achieving high-fidelity surrogates of 3D RANS simulations of axial compressors. The findings highlight the potential for AI to significantly enhance design efficiency and innovation in turbomachinery simulations.
Pulga et al. (2025) studied this question.