Structural Optimization of Turbomachinery Rotors: A Study of Machine Learning Surrogate Models

The use of machine learning models as surrogates for turbomachinery component design has demonstrated promising results in various previous studies. This includes a previous study by the authors, where Finite Element Analysis (FEA) data was used to train a neural network model, which was then used for optimizing rotor blade design to reduce stress without altering blade thickness. However, many decisions affect the quality and accuracy of the optimization and the effects of these options have not been rigorously tested. This study investigates key factors influencing this process: the selection of surrogate models, the number of design variables, and the training sample size. In this study, the design space is sampled using Latin Hypercube Sampling (LHS). Samples of various sizes ranging from 500 to 5000 are used to train the surrogate models. Two machine learning models from the Python package scikit-learn are evaluated: Random Forest Regressor (RF) and Multi-layer Perceptron Regressor (MLP). The MLP, a neural network, is further analyzed by varying the number of hidden layers from 1 to 5, with each layer containing 100 neurons. Due to the stochastic nature of these models, each is assessed using 100 different random initializations. Model accuracy is first evaluated using a validation set, generated independently from the training set using a separate LHS. This validation set, which is 20% the size of the training data, is also used for early stopping in the neural network models. The accuracy is further evaluated on a separate test set that is never used during any part of the model training process. Finally, the best-performing surrogate model from each type is employed for design optimization using a differential evolution optimization tool from SciPy. The results of the surrogate model optimization are evaluated using FEA to confirm the predicted performance.