Efficient ML-Based Prediction of Turbomachinery Blade Performance With B-Spline Surface Representation

Abstract In turbomachinery blade design, rapid and accurate performance prediction is essential to accelerate optimization and reduce reliance on costly high-fidelity simulations. Traditional data-driven approaches often use dense surface point-cloud representations as input features, requiring extensive training datasets and computational resources. This work presents a more efficient methodology leveraging a compact B-Spline-based surface representation, where control points serve as input features, significantly reducing geometric dimensionality and computational overhead. A systematic Design of Experiments (DoE) is performed to generate a diverse set of blade geometries for NASA Rotor 67. Each design is evaluated via computational fluid dynamics (CFD) simulations in ANSYS CFX, providing key aerodynamic performance metrics such as isentropic efficiency. We train and compare Graph Convolutional Neural Networks (GCNN) and Random Forest Regression (RFR) models to predict blade performance directly from the reduced control-point parameterization. Incorporating first- and second-order geometric derivatives (gradients and Laplacians) into the feature set significantly enhances predictive accuracy and stability, capturing essential curvature-related flow physics. Results demonstrate that this B-Spline-based, CAD-centric methodology can achieve competitive accuracy in predictions with as few as 150–200 training simulations—comparable to other GCNN-based approaches. Consequently, the proposed framework reduces training overhead from days to minutes, enabling faster, more cost-effective turbomachinery design workflows and guiding optimization toward high-performing blade geometries.