This thesis proposes a surrogate model based on feed-forward neural networks for the rapid generation of optimal guess solutions for low-thrust transfers. The use of low-thrust propulsion in interplanetary missions, increasingly adopted for its high propulsive efficiency, calls for new methods for the preliminary design of trajectories. At this stage, a large number of mission configurations and launch windows must be explored. However, generating each optimal solution requires solving a nonlinear optimization problem, which has a high computational cost and is not sustainable for extensive trade-off analyses. Machine learning-based surrogate models are one of the most promising strategies for significantly facilitating the task.