In this study, we tackle the problem of searching for the most favourable pattern of link capacities allocation that makes a power transmission network resilient to cascading failures with limited investment costs. This problem is formulated within a combinatorial multi-objective optimization framework and tackled by evolutionary algorithms. Two different models of increasing complexity are used to simulate cascading failures in a network and to quantify its resilience: a complex network model (namely, the Motter-Lai (ML) model) and a more detailed and computationally demanding power flow model (namely, the ORNL-Pserc-Alaska (OPA) model). Both models are tested and compared on a case study involving the 400kV French power transmission network. The results show that cascade-resilient networks tend to have a non-linear capacity-load relation: in particular, heavily loaded components have smaller unoccupied portions of capacity, whereas lightly loaded links present larger unoccupied portions of capacity (which is in contrast with the linear capacity-load relation hypothesized in previous works of literature). Most importantly, the optimal solutions obtained using the ML and OPA models exhibit consistent characteristics in terms of phrase transitions in the Pareto fronts and link capacity allocation patterns. These results provide incentive for the use of computationally-cheap network-centric models for the optimization of cascade-resilient power network systems, given the advantages of their simplicity and scalability.