Abstract: 3D prospecting prediction has become an indispensable technique for deep and peripheral ore exploration in current mines, where the reliability of such predictions is heavily dependent on the 3D characterization of ore-controlling structural surfaces. Traditional methods rely on manual extraction of morphological indicators （e.g., distance, curvature, and slope）, which are highly subjective and difficult to fully characterize complex 3D spatial ore-controlling patterns. To address this issue, by taking the Xiadian Gold Deposit as the research object, a 3D intelligent prospecting prediction model based on the combination of Laplace−Beltrami feature descriptors and convolutional neural networks was proposed. This method leveraged the Laplace−Beltrami features to objectively and quantitatively characterize the 3D geometry of faults and orebodies. Unstructured 3D geometric information was converted into 2D multi-channel images using the feature projection method, which were then input into the CNN model for deep learning and adaptive feature fusion. The results demonstrate that the proposed model can effectively identify deep concealed mineralization spaces. Four high-potential prospecting targets have been delineated in the deep part of the Xiadian Gold Deposit, providing a scientific basis for the subsequent ore exploration work of the mine.