Abstract: A reduced-order model for solving three-dimensional flow fields in cascades was proposed, which reduced RANS equations to corresponding Euler equations coupled with viscous correction source terms. The inviscid base three-dimensional flow field was obtained by solving Euler equations. A multilayer perceptron neural network was then employed to predict the viscous source terms under various operating conditions and geometric parameters, thereby correcting the Euler solutions. The model was applied to solve the flow fields of Rotor 37 and a 1.5-stage Aachen turbine. Results show that the machine learning correction method, without requiring extensive training samples, improves the accuracy of the Euler solutions by an order of magnitude, accurately capturing the three-dimensional flow characteristics within the cascade. This approach holds significant application value for reducing hardware requirements and enhancing computational speed for cascade optimization computations and throughflow design of turbomachinery.