This paper deals with designing of cryogenic distillation systems for the separation of hydrogen isotopes in a thermonuclear fusion process. The design must minimize the tritium inventory in the distillation columns and satisfy the separation requirements. This induces the optimization of both the structure and the operating conditions of the columns. Such a problem is solved by use of a Mixed-Integer NonLinear Programming (MINLP) tool coupled to a process simulator. The MINLP procedure is based on the iterative and alternative treatment of two subproblems : a NLP problem which is solved by a reduced-gradient method, and a MILP problem, solved with a Branch and Bound method coupled to a simplexe. The formulation of the problem and the choice of an appropriate superstructure are here detailed, and results are finally presented, concerning the optimal design of a specific isotope separation system.