Nowadays, Machine Learning techniques offer fast and efficient solutions for classification problems that would require intensive computational resources via traditional methods. We examine the use of a supervised Random Forest to classify stars in simulated galaxy clusters after subtracting the member galaxies. These dynamically different components are interpreted as the individual properties of the stars in the Brightest Cluster Galaxy (BCG) and IntraCluster Light (ICL). We employ matched stellar catalogues (built from the different dynamical properties of BCG and ICL) of 29 simulated clusters from the DIANOGA set to train and test the classifier. The input features are cluster mass, normalized particle cluster-centric distance, and rest-frame velocity. The model is found to correctly identify most of the stars, while the larger errors are exhibited at the BCG outskirt, where the differences between the physical properties of the two components are less obvious. We investigate the robustness of the classifier to numerical resolution, redshift dependence (up to z = 1), and included astrophysical models. We claim that our classifier provides consistent results in simulations for z < 1, at different resolution levels and with significantly different subgrid models. The phase-space structure is examined to assess whether the general properties of the stellar components are recovered: (i) the transition radius between BCG-dominated and ICL-dominated region is identified at 0.04 R200; (ii) the BCG outskirt (>0.1 R200) is significantly affected by uncertainties in the classification process. In conclusion, this work suggests the importance of employing Machine Learning to speed up a computationally expensive classification in simulations.