Rapid Algorithmic Blade Design Applying Machine Learning From Shape Optimization to Satisfy Multidisciplinary Constraints

Abstract This paper presents a new, rapid, flexible approach for turbine blade section design combining algorithmic and inverse techniques to enable automatic generation of blade sections guaranteed to conform to multidisciplinary requirements, with the aim of accelerating turbomachinery design iterations. The approach links a base algorithm to parametrize and generate blade section geometry conforming to structural and manufacturability constraints such as section area, trailing edge radius, and exit wedge angle with a 2D CFD solver to calculate surface isentropic Mach number profile for aerodynamic performance evaluation. To achieve blade sections with smooth surface and lenticular Mach number profile concave up on the pressure-side and concave down on the suction-side, the base algorithm is tuned by a surrogate inverse model trained by machine learning from pre-generated tuning data obtained by case-by-case shape optimization for a range of design conditions. A weighted objective function is applied to quantify both geometric and aerodynamic quality of blade sections for the optimization. Shape optimization improves output section quality by 54–87% compared to the untuned algorithm. Aerodynamically, suction-side flow separation is eliminated in the optimized sections, giving 70% less pressure loss compared to the untuned algorithm for the best cases. Across all conditions spanning the examined design space, the surrogate model successfully captures most of this improvement, yielding blade sections of similar quality to explicit optimization sufficient to meet the geometric and aerodynamic requirements for design. Furthermore, section quality is preserved even if imposed structural and manufacturability constraints are perturbed within typical margins, guaranteeing blade sections that are always viable for practical use. Blade sections from the surrogate-tuned algorithm are output within minutes, eliminating time-intensiveness of existing manual or case-by-case design approaches.