Annotation: The paper addresses the problem of 3D-representations and automatic synthesis of gas turbine blades shapes. First, we implemented a parametric method of descriptor-based representation using Bernstein polynomials and generalized it to produce controllable 3D-shapes. Then, we proposed a method of automatic synthesis of 3D-shapes based on the use of generative ML models for aerodynamic profiles. This method helps to reduce the number of geometric design variables used in the optimization of the aerodynamic shape of blades. Moreover, it enables automatic synthesis of 3D-shapes with representation independent of shapes level of detail. Its implementation is based on generative-adversarial network BézierGAN and makes it possible to produce arbitrary sized datasets of 3D blades having aerodynamic shapes. Finally, by interpreting and visualizing the generator’s latent space, we observed the subset of latent variables that has the most importance for rapid prototyping of gas turbine blades

Keywords: gas turbine blade, dataset, 3D object representation, machine learning, generative-adversarial network, Bézier curves, Bernstein polynomials