Mathematical geological models are being increasingly used by natural resources delineation and planning agencies for mapping areas of mineral potential in order to optimize land use in accordance with socio-economic needs of the society. However, a key problem in spatial-mathematical-model-based mineral potential mapping is the selection of appropriate functions that can effectively approximate the complex relationship between target mineral deposits and recognition criteria. This research evaluates a series of mathematical geological models based on different linear and non-linear functions by applying them to base-metal potential mapping of the Aravalli province, western India, where several significant base-metal deposits are already known. Linear models applied in this research are an extended weights-of-evidence model and a hybrid fuzzy weights-of-evidence model, while non-linear models are knowledge and data-driven fuzzy models, a neural network model, a hybrid neuro-fuzzy model and an augmented naive Bayesian classifier model. From a conceptual model of base-metal metallogenesis in the study area, host rock lithology, stratigraphic position, (palaeo-)sedimentary environment, association of mafic volcanic rocks and proximity to favorable structures are identified as recognition criteria for base-metal mineralization and, in the form of predictor maps, constitute the input for the models. The parameters of the knowledge-driven fuzzy model are estimated from the expert knowledge, while those of the neural network and Bayesian classifier model are estimated from the data. The two hybrid models use both expert knowledge and data for parameter estimation. As compared to the linear models, the non-linear models generally perform better in predicting the known base-metal deposits in the study area, including the known deposits that are characterized by unusual geological settings. This is attributed to the ability of non-linear functions to (a) better approximate the relation between mineral deposits and recognition criteria and (b) recognize and account for possible dependencies amongst recognition criteria. Although the linear models do not fit the data as efficiently as the non-linear models, they are easier to impleme nt using basic GIS functionalities and their parameters are more amenable to geoscientific interpretation. In addition, the linear models are less susceptible to the curse of dimensionality as compared to non-linear models, which makes them more suitable for applications to mineral potential mapping of the areas where there is a paucity of training mineral deposits. The hybrid models that conjunctively use both knowledge and data for parameter estimation generally perform better than purely knowledge-driven or purely data-driven models. This is attributed to the capability of the hybrid models to cross-compensate the deficiencies in either the knowledge-base or in the database. The output of various mathematical geological models indicate that about 10% of the study area has potential for base-metal deposits. The high-favorability zones tend to reflect a strong lithostratigraphic control on base-metal mineralization in the study area and the importance of crustal-scale faults in spatial localization of base-metal mineralizations in the study area. The low proportion of base-metal potential area delineated and the high prediction rates, which vary from 83% to 96% depending on the model applied, indicate not only (a) the efficiency of the mathematical geological models in capturing the complex relationship between the target mineral deposits and the deposit recognition criteria, but also (b) usefulness of the favorability maps as inputs to natural resources planning for optimizing land use in the study area.