Most machine learning monitored algorithms used in the preparation of mineral potential prediction maps, require noise-free data to achieve high-performance and reliable results. To discover the hidden structures of a data set, unsupervised clustering methods are highly efficient. Therefore, in this paper, a combination of supervised and unsupervised methods was attempted using training and testing data to prepare a potential map with high accuracy in the Sonajil copper-gold deposit, located in NW of Iran. Here a semi-supervised Bayesian algorithm utilized to map the mineral landscape. At First, raster layers of 10 exploratory features were prepared. Then, 27 exploratory boreholes drilled were divided into 4 classes C1 to C4 based on the copper concentration, and from each class, two boreholes were selected and 100-meter buffering was performed around these boreholes to extract 1113 training data. Then, the existing data were clustered by the FCM method and the whole data and clustering data were entered into the Bayesian algorithm to evaluate the accuracy of Bayesian classifier method in 5 different clusters. The results show an increase in the average accuracy while using the clustered data instead of whole data for MPM mapping. Also, in case of not using cluster 5 data, it has the highest accuracy of 96%. To validate the Bayesian semi-supervised method, boreholes that were not used in data training were employed. Overall results based on the classification of boreholes and prepared potential map indicate accuracy of the Bayesian semi-supervised method as an optimal result for further exploration and delimiting more mineralization targets.