Session: 36-02 Surrogate Model-based applications

Paper Number: 153577

Artificial Intelligence Surrogates for Flow Modeling of an Axial Compressor 

Development of 3-dimensional airfoil shapes for turbomachines is an iterative process with significant dependency on computational fluid dynamics (CFD). During the design phase, Design Engineers (DE) have to run a new Computational Fluid Dynamics (CFD) simulation for each design iteration to compute the new performance. Each CFD simulation involves creation of quality mesh (fine-tuned by hand) and multistage computations (at least 10 for a speedline), making the process very time consuming and requires considerable human and computational resources.  This limits the number of the design interactions in the development phase, often leading to suboptimal airfoil designs satisfying the minimum product requirements.  

With Artificial Intelligence (AI), we can speed up and improve such design. The DE will still have to create meshes and CFD simulations, but only for the initial geometry set. Then, a surrogate model for the flow field will be trained. Any subsequent geometry change will be assessed by applying the surrogate within minutes, and a complete flow field will be predicted. Unlike other approaches where only a selected KPIs are predicted with AI, our approach allows the designer to investigate the flow for any insights they may be interested in, without having to specify them beforehand. Quality mesh will not be needed anymore, only a geometry file. Also, AI can be used for uncertainty quantification, sensitivity analysis, generally all those tasks that require multiple queries to the model. This way, DE will be able to investigate new geometries very efficiently or to find optimal geometries given the product leader requirements. But turbomachinery simulations need a very high level of accuracy.

We investigate multiple different ways to build these surrogates, using open source libraries. First, we investigate two different approaches to represent the 3D geometry: in one, we compute a 3D occupancy mask for the geometry, and then extract a latent representation of it, while in the other we directly compute a latent representation of the CFD mesh. Then, in both cases we train regression models to predict the flow field quantities at each point in space, given the geometry representation and operating conditions. Finally, we also test an approach where we only consider the surfaces defining the CFD simulation, and we train an advanced geometric deep learning model to predict the flow quantities on such surfaces. We apply these methods to axial compressors design using legacy CFD simulations, and we then apply our post-processing tools to compare ground truth with predictions.  We find that, by selecting the right models and tuning hyperparameters appropriately, we can create high-fidelity surrogates of 3D RANS simulations of axial compressors.

We also assess the business value of these methods, by investigating the tradeoff between the size of the training dataset and the performance of the best model we can train on it without overfitting. This is especially important since building the training dataset incurs a high computational and human time cost for the business, that in recent literature has been sometimes overlooked, but it is of paramount importance for the industrial application of these methods.

Presenting Author: Leonardo Pulga Baker Hughes

Presenting Author Biography: Leonardo Pulga is a Senior Data Scientist in Baker Hughes, based in Florence. He holds a Ph.D. in Engineering from the University of Bologna on the interaction between artificial intelligence and CFD simulations in the field of energy systems. His main work is related to the development of surrogate models for design exploration and optimization in collaboration with engineering design teams.

Authors:

Leonardo Pulga Baker Hughes
Andrea Panizza Baker Hughes
Laure Barrière Baker Hughes
Giovanni De Martino Baker Hughes
Vittorio Michelassi Baker Hughes
Corrado Burberi Baker Hughes
Satish Koyyalamudi Baker Hughes