Abstract. Accurate and rapid assessment of turbulent flows in turbine working channels remains one of the key challenges in computational gas and fluid dynamics, especially under conditions of high demands on the efficiency and reliability of turbomachinery. Traditional approaches, such as Reynolds-based simulation (RANS) or large eddy simulation (LES), although providing acceptable accuracy, are associated with high computational resource and time costs. This paper proposes an alternative approach based on the use of convolutional neural networks (CNN) as a surrogate model for reproducing three-dimensional velocity and pressure fields in turbulent flows. The developed architecture is based on a modified version of U-Net and adapted for three-dimensional input data. A comparison with LES results showed that the proposed model is capable of reproducing key flow characteristics, including vortex structure and pressure gradients, with a high degree of accuracy. At the same time, a significant acceleration of calculations is achieved — up to 10³ times compared to classical numerical methods. The proposed neural network model demonstrates stability to changes in geometric parameters and can be easily reconfigured for other channel configurations. The results obtained highlight the potential of deep learning in turbulent flow modelling and open up prospects for the integration of such models into real-time engineering calculations.

Keywords: turbulent flow modeling, convolutional neural networks (CNN), U-Net architecture, surrogate modeling, Large Eddy Simulation (LES), turbomachinery, data-driven computational fluid dynamics (CFD).