The radial inflow turbine is the core component of the organic Rankine cycle, and the application of splitter blades is an effective way to reduce its flow loss. It is helpful to study the energy loss characteristics of radial turbines with splitter blades and the optimal geometrical parameters of splitter blades for further improving their aerodynamic performance. In this paper, an optimization framework based on the combination of back-propagation neural network and non-dominated sorting genetic algorithm II is developed. Considering the turbine efficiency as the objective, the geometric parameters of splitter blades, including relative length of the blade, relative distance of the leading edge, circumferential offset, and outlet angle of the blade, are optimized. In addition, the entropy production method is used to locate and quantitatively evaluate the energy loss of the radial turbine with the optimal splitter blades. The results indicate that the genetic algorithm optimized back-propagation neural network model can accurately predict turbine efficiency with a maximum deviation of 2.47 %. In addition, the optimal geometrical parameters of the splitter blade are related to the turbine geometry. In this case, the optimal splitter blade has a relative blade length of 0.639, a relative leading edge distance of 0.3, a circumferential offset of 0.442, and an outlet angle of 38°. Compared with the original turbine, the efficiency of the radial turbine with optimal splitter blades increases by 4.65 %. The total entropy production of rotor and turbine is reduced by 33.3 % and 44.1 % respectively by the installation of optimal splitter blades. Compared with the case without splitter blades, the turbine with the optimal splitter blades can achieve higher efficiency at a flow rate 40 % higher than the design value, and it expands the efficient working range of the turbine by 60 %. This work can provide guidance for the optimization design of radial inflow turbines.