This paper proposes a machine learning framework for efficient cracks localization and characterization in configurations of fastener inspections using surface eddy current probes. The learning algorithm relies on a supervised learning procedure. In this work, the training set is a simulated set of input/output couples. Furthermore, in order to assess variability associated to real inspection procedures, uncertainties impacting probe lift-off and tilt have been introduced in the inversion scheme in order to evaluate its robustness. Results obtained through a numerical validation campaign shown that both cracks localization and characterization are possible even when some uncertainties are taken into account. Impacts of uncertainties on the prediction accuracy as well as on the CPU time efficiency are finally discussed.