A linearized model of active magnetic bearings (AMBs) is normally used for determining the settings of decentralized PID controllers, which are used in industrial applications. Parameters of the linearized model, i.e., current gain and position stiffness, are defined for the nominal operating point and, thus, vary according to the operating point. These variations are analyzed by the finite element (FE) computation. The obtained results show considerable variations for the discussed parameters. Therefore, AMB geometry is numerically optimized, using differential evolution in combination with FE computation. The optimization objective is a minimal variation of the current gain, and position stiffness. The optimized AMBs shows good properties according to the linearized model over the entire operating range. In this way, the robustness of the closed-loop controlled AMB system is increased.