During the last decade, organic Rankine cycle (ORC) turbogenerators have become very attractive for the exploitation of low-temperature heat sources in the small to medium power range. Organic Rankine cycles usually operate in thermodynamic regions characterized by high pressure ratios and strong real-gas effects in the flow expansion, therefore requiring a nonstandard turbomachinery design. In this context, due to the lack of experience, a promising approach for the design can be based on the intensive use of computational fluid dynamics (CFD) and optimization procedures to investigate a wide range of possible configurations. In this work, an advanced global optimization strategy is coupled with a state-of-the-art CFD solver in order to assist in the design of ORC turbines. In particular, a metamodel assisted genetic algorithm, based on the so-called `off-line trained’ metamodel technique, has been employed. The numerical solutions of the two-dimensional (2D) Euler equations are computed with the in-house built code zFlow. The working fluid is toluene, whose thermodynamic properties are evaluated by an accurate equation of state, available in FluidProp. The computational grids created during the optimization process have been generated through a fully automated 2D unstructured mesh algorithm based on the advancing-Delaunnay strategy. The capability of this procedure is demonstrated by improving the design of an existing one-stage impulse radial turbine, where a strong shock appears in the stator channel due to the high expansion ratio. The goal of the optimization is to minimize the total pressure losses and to obtain a uniform axisymmetric stream at the stator discharge section, in terms of both the velocity magnitude and direction of the flow.