Railway dynamics simulations are increasingly used to characterise system behaviour, whether for design, certification, monitoring or maintenance purposes. However the system is complex due to the presence of strong non-linearities and sources of variability (friction coefficient at the wheel/rail contact, wind, etc.), and the challenges remain numerous. In this context, this work proposes a machine-learning based method to optimize the suspension elements of an innovative train. Using rigid-body modelings of the train and the track, the objective is to find the train parameters maximizing its stability while limiting its aggressiveness towards the infrastructure. The problem is solved in a Bayesian framework.