Along with the developing technology, robotic systems have started to take place in areas where there is one-to-one interaction with people, as well as their use in industrial areas. As the robotic system began to take place in daily life, safety and reliability between humans and robots have become a critical issue. In this context, a series elastic actuator has been developed for the aforementioned robotic systems, which has an elastic element placed in series between the motor output and the mechanical output. In this thesis, the torsion spring, as a critical part for the rotary series elastic actuators of rehabilitation robots, which helps support the extension and flexion of the knee joint during physical therapy of individuals with lower extremity disorders, is discussed. First of all, the data required for modeling was produced by making analyses with the design of experiment and finite element method. In line with the design goal of a light, compact, durable and stiff spring, the torsion spring whose topology was determined was modelled using a hybrid method: Neuro-regression approach and cross-validation technique. To minimize the mass and von Mises stress of the torsion spring, the thickness of the spring and the inner corner radius of the flexible leg are taken as the design variables and multi-objective optimization problems are created. The design and optimization of the torsion spring was done with the help of Differential Evolution, Nelder-Mead, Random Search and Simulated Annealing algorithms. By comparing the obtained optimization results with the finite element method and the results in the literature, it has been seen that the model and optimization methods used in the study are reliable and applicable.