Application of Deep Machine Learning Methods for Designing Turbomachinery Blade Cascades

Abstract The article presents an integrated procedure for 3D designing of turbomachinery stages, which combines the use of a blade passage shape parametric description and neural networks. The procedure has at its heart a blade parametric model on the basis of Bezier curves, which employs eight to ten controlling parameters (thickness, edge radiuses, inlet/outlet angles, etc.), due to which it becomes possible to produce smooth profiles and reduce the problem dimensionality. By using such approach, a fully connected neural network for predicting the profile loss coefficient in nozzle vane cascades has been developed. The network architecture includes three hidden layers with Swish activation functions and Dropout regularization. The network was trained on a dataset of 2D CFD computations covering a wide range of geometric parameters, Mach numbers (0.6–0.95), and Reynolds numbers (1.6 × 105–1.1 × 106). The trained model’s median prediction error amounted to less than 0.24%. The procedure was applied for designing the steam turbine last stage. Nine nozzle vane sections have been designed and optimized, and the 3D cascade has been shaped. The verification carried out using 3D CFD simulation in the ANSYS CFX environment has confirmed the approach adequacy: the loss distribution pattern over the nozzle vane cascade height is in qualitative agreement with the calculation results. Comparison carried out for middle sections has shown satisfactory quantitative coincidence (the average loss predicted using the neural network was equal to 4.24% against 4.83% estimated from CFD simulation). In addition, the problem of improving the erosion resistance has been solved. For the near-tip section at the nozzle vane cascade relative height equal to 0.75, its separation performance has been optimized. By using the second neural network, which predicts the droplet motion trajectories and points at which they collide with the surface, the leading edge radius and the suction and pressure side curvatures were varied. The aim of the optimization was to enhance the moisture deposition on the profile surface for its subsequent controlled removal through the moisture removing slit, which would help decrease the rotor blade erosion wear. It is shown that by integrating the parametric models and neural network, it is possible to develop a tool for effectively designing turbomachinery stages, predicting the aerodynamic characteristics, and solving a problem connected with two-phase flows.