Abstract
A Machine Learning approach is applied to the open domain test case Rotor 37 to predict the performance of a family of turbomachinery. A deep learning approach implemented in the form of a geometric convolutional neural network is used. A first low geometrical variability parametric representation of the Rotor 37 is created to build training dataset from which the machine learning model is trained. After the training, the model predicts machine performance parameters in a few seconds. Then, this AI model is used in an optimization process to improve Rotor 37 performance and an increase in isentropic efficiency of 2.86% is obtained. A CFD validation is performed to confirm the improvement with an increase of 1.81% in isentropic efficiency. However, the achievable performance increase is limited by the parametric representation generated, given its low degree of freedom. Therefore, a new parametric representation is generated allowing more complex variations of the Rotor 37 design. It is demonstrated that a continuous learning approach, which is used to extend the model can successfully be applied to extend the trained AI model to the larger geometrical variability at significantly reduced model training costs. Then the optimization procedure is repeated and an increase of isentropic efficiency of 3.25% is obtained. A CFD simulation confirms an improvement of the performance of 2.35% in isentropic efficiency.
