Turbine is one of the key components of ocean thermal energy conversion system, and its aerodynamic performance and geometric structure directly affect the performance of the system. At present, the design methods of radial inflow turbines can be roughly divided into three categories: trial method, best velocity ratio method and screening method, which follow no concrete rules and are not comprehensive, and rely mostly on the designers' experience. This study proposes a fully data-based non-parametric model identification and optimization method for ocean thermal energy conversion radial inflow turbines is proposed, and takes the 25 kW R134 working fluid turbine of the ocean thermal energy conversion system as an example for verification and analysis. First, the optimal Latin hypercube sampling method was used to complete the experimental design with seven initial thermodynamic parameters as independent variables. And then according to the test data and the Gaussian process regression identification, the non-parametric turbine thermodynamic model was obtained. At last, on the basis of this model, the NSGA-III multi-objective optimization algorithm was used to optimize the design with the goal to achieve maximum efficiency and minimum diameter, and the optimized results were obtained. The results show that the non-parametric model based on the Gaussian process proposed in this study has higher accuracy, and the optimization results obtained based on this model are more efficient and have a more compact structure. The research work described in this study saves time and cost for completing the aerodynamic performance design of the turbine and further improving the expansion efficiency of the radial inflow turbine. It is of great significance for the practical application of promoting ocean thermal energy conversion.