Hybrid Surrogate Modeling Approach for Data Reduction and Design Space Exploration of Turbine Blades

The conventional iterative optimization of turbine blades using computer-aided engineering simulations is resource-intensive, with high costs and time demands, as well as significant challenges in computational requirements and data management. The three-dimensional (3D) simulation data generated from computational fluid dynamics (CFD) and finite element analysis (FEA) for various blade geometries can range from hundreds of gigabytes to multiple terabytes, complicating long-term storage and access. To address this, we propose a machine learning-based methodology for data reduction and prediction of 3D surface field data. Our approach involves developing a convolutional variational auto-encoder (VAE), consisting of an encoder and a decoder. The encoder compresses the input data into a representation of reduced dimensionality in a latent space, while the decoder reconstructs the data from this latent space back to its original form. This significantly reduces the amount of stored data, facilitating long-term use. Additionally, we train a fully connected feedforward multilayer perceptron (MLP) to map geometry parameters, which generate blade variations, to the latent space. By combining the MLP with the VAE's trained decoder, we create our proposed multilayer perceptron–variational auto-encoder (MLP–VAE) hybrid model capable of predicting surface field data for new, unseen blade geometries. The MLP–VAE generates latent representations and surface field results with high accuracy (>97%) and without additional computational costs, offering a highly efficient and scalable solution for turbine blade optimization.