Abstract: Deep learning has been widely applied in the field of mineral exploration. However, complex network models generally involve significant computational demands. To enhance the accuracy and efficiency of intelligent mineral exploration predictions, this study addressed the shortcomings of traditional convolutional neural networks （CNNs） in modelling geological feature correlations for mineral resource prediction, as well as the low data fusion efficiency of complex deep learning architectures. Taking the Yundukala Area in Xinjiang as the study area, the study integrated geological, aeromagnetic, gravity, and geochemical data and proposed an intelligent mineral exploration prediction method based on the convolutional mixer （ConvMixer） model, aiming to achieve more accurate regional prediction results using a simpler architecture. Results indicate that ConvMixer can efficiently perform spatial fusion of multi-source data and extraction of mineralization features. After optimizing parameters, compared with that by CNN, the proportion of predicted area covered by ConvMixer decreases from 21.18 % to 13.82 %, with a model accuracy of 98.3 %, and the predicted target areas are more contiguous and complete; compared with that for Transformer, the prediction time per round for ConvMixer is reduced from 2 s to 1 s, improving computational efficiency. Based on this model, two mineral exploration prediction zones are delineated, both located within core mineral-bearing strata and characterized by well-developed NW-trending faults and geophysical anomalies, indicating favorable mineralization conditions. ConvMixer combines the advantages of high accuracy and high efficiency, providing an efficient and feasible technical solution for intelligent cobalt deposit resource exploration prediction in regions with massive amounts of grid-based geological data.