Starting from a primary perceptual evaluation of a set of car dashboards, we propose to build a Bayesian network (BN) between perceptual attributes and design attributes. Two types of learning processes may be considered: supervised BN when the prediction on a targeted attribute must be optimised and unsupervised BN otherwise. These two types of BNs are considered along three design simulation scenarios: the direct scenario which consists of the prediction of a design change impact on customer perceptions, the inverse scenario for fixing design characteristics so as to result in an expected customer perception, and a more realistic combined scenario.