The Carajás Mineral Province (CMP) located in the southeast of the Amazon craton, northern Brazil is one of the most important provinces of the world hosting a high number of large-tonnage Cu–Au deposits. This potential is due to overprinted mineralization events dated from the Neoarchean to Paleoproterozoic, including VMS, IOCG, and Cu–Au granite-related systems. The discovery of a new mineral deposit is linked to direct detection methods that have an elevated cost. Mineral potential modeling is a multi-source geoscience data that can help outline high prospective areas, containing new mineral exploration targets in the order to reduce the exploration risk. This work employed one knowledge-based method, multi-class index overlay, and one machine-learning algorithm, random forest, to produce two mineral potential maps in a district scale for Cu–Au deposits in the easternmost CMP (Parauapebas area) using a mineral system-based approach. This approach consists of translating the critical ore-forming processes of the mineral system components (source of fluids, metals, and ligands, source of energy, pathways, ore deposition gradient) into spatial data (evidence maps). For this study, ten evidence maps were created using the Spatial Data Analyst in GIS, then for the multi-class index overlay, each evidence map received a weight according to its relevance in the modeling, and integrated through the Image Analyst Tools. For the random forest algorithm, meanwhile, the model was trained four times using the Orange Data Mining software. The performance of the models was measured through the classification of accuracy, ROC curve, AUC values, predictive and success-rate curves. Both models achieve good performance with classification accuracy and AUC values of 80% and 0,805 for the multi-class index overlay and values of 76,4% and 0,812 for the random forest. Although the models have similar performance, the analysis of the curves showed that the random forest model was able to predict 82% of the locations of the known deposits, selecting about 25% of the total study area, while the multi-class index overlay predicted 72% of the location of the known deposits over 35% of the study area. Furthermore, both models predicted less than 10% of the study area as highly prospective helping prioritize areas for research and follow-up copper and gold exploration activities.