Turbomachinery remains a cornerstone technology for aviation, power generation, and industrial energy conversion. The aerodynamic design of turbine cascades is particularly challenging in the transonic regime, where locally supersonic regions, shock waves, and shock-boundary-layer interactions strongly affect performance and frequently lead to non-convergent simulations or physically inconsistent designs. These difficulties make brute-force design space exploration impractical even with fast two-dimensional solvers. This thesis develops a data-driven pipeline for rapid transonic 2D turbine cascades design and optimization. A neural network surrogate model is trained on a database of inviscid-viscous coupled simulations performed with MISES, learning to predict aerody- namic performance from blade geometry across a wide range of operating conditions defined by the duty vector (α1, α2, M2, Re) and the trailing edge thickness tT E as an additional design parameter.