This paper aims to present a data-driven method for aerodynamic design of cascade thrust reversers installed on aircraft turbofan engines. In the new method, thrust reverser cascades can be optimally and efficiently designed, offering excellent thrust reverse efficiency and flow performance.
The method involves three steps: optimized Latin hypercube sampling (OLHS), obtaining a data set through computational fluid dynamics (CFD) and predicting and optimizing designs using deep neural networks (DNN). The OLHS evolved from the traditional Latin hypercube algorithm, integrating the maximin distance criterion and particle swarm optimization. The quasi-three-dimensional vane flow simulations were carried out to construct a data set for predicting the aerodynamic performance of cascades. A fully connected DNN with four hidden layers was well trained. By using appropriate aerodynamic performance evaluation metrics as the objective functions, the optimal thrust reverser cascades could be found.
Hyperparameters had a significant impact on the prediction accuracy of the DNN model. After comparative studies, the number of hidden layers was set to be 4, the number of artificial neurons of each hidden layer was set to be 256 and the learning rate was determined to be 0.0002. In the test set, most of the relative errors are within ±3%. The prediction of thrust efficiencies and flow coefficients of 10,000 sample points using the DNN was completed in less than 1 min.
Before this, the data-driven approach had not been applied to the aerodynamic design of cascade thrust reversers. In comparison to traditional experimental and computational methods, the data-driven method established in this paper assists designers in more rapidly acquiring aerodynamic performance and more comprehensively uncovering physical principles. Simultaneously, the design solutions recommended by this method can also serve as a significant reference for other approaches, rendering the aerodynamic design of the cascade thrust reverser more robust.
