The Gaia GraL catalogue of clusters consist of 2,058,962 clusters with three components and 70,697 clusters with four components. Each of these clusters satisfied the following conditions: i) Clusters are composed of three or four images in order to provide a sufficient number of constraints for identifying gravitational lens candidates. ii) Their constituent images have negligible parallaxes, plx, and proper motions, (pmra', pmdec) where pmra' = pmra * cos(dec). Specifically, we required that plx-3*e_plx<4mas and abs(pm)-3*epm<4mas (where eX is the mean error on X and pm stands for pmra' and pmdec). iii) The maximal angular separation between any pair of images is below or equal to 6 arcsec. iv) The absolute difference in G magnitude between components is lower or equal to 4mag. v) Clusters are located in regions with a mean field density lower than 60000 objects/deg2. The mean density of objects is computed within a radius of 30 arcsec around each cluster. Based on a supervised learning method, called extremely randomized trees (ERT), each cluster is assigned a discriminant value: the ERT probability. This ERT probability reflects the ability of the clusters to be matched to the image positions and relative magnitudes produced from simulations of lens systems based on a non-singular isothermal ellipsoid lens model (Kormann+, 1994) in the presence of an external shear (Kovner, 1987). These probabilities do not constitute probabilities in a mathematical sense, although they can be translated into expected ratios of identification of gravitational lenses (the true positive rate, TPR) and to expected ratios of misclassification of groups of stars as gravitational lenses (the false positive rate, FPR) through the use of an appropriate cross-validation procedure (see the description of the roc*.dat files). As we expect some of the lensed images to be missing from Gaia DR2, all combinations of three and four images were considered fo building the ERT. The resulting ERT models wil be referred to as ABCD, ABC, ABD, ACD and BCD where A, B, C, D identify the images we used for building the corresponding ERT model, assuming these are sorted in ascending order of G magnitude. Finally, the receiver operating characteristics (ROC) curves from files roc.dat, rocabc.dat, rocabd.dat, rocacd.dat, and rocbcd.dat are obtained by computing the TPR and FPR that are associated with all ERT probabilities coming from the corresponding ERT model based on a test set and a validation set of observations. The test set of observations (TS) is composed of 5*107 gravitational lenses simulated from a non-singular isothermal ellipsoid lens model in the presence of an external shear and of the same number of simulated contaminants where the relative image positions and fluxes are randomly drawn from a uniform distribution. The validation set of observations (VS) is composed of 12 known lensed systems having four detections in Gaia DR2 (2MASSJ11344050-2103230, J1606-2333, WGD2038-4008, HE0435-1223, SDSS1004+4112, PG1115+080, B1422+231, 2MASXJ01471020+4630433, 2MASSJ13102005-1714579, J1721+8842, WFI2033-4723 and RXJ1131-1231) and of 1e6 clusters we randomly extracted from Gaia DR2 with a size smaller than 30 arcsec and a maximal absolute difference in G magnitude between their constituent images <4mag. (6 data files).< Réduire