Pyrite is a ubiquitous metal sulfide in gold mines, and its trace element content information can reveal the composition and crystallization characteristics of related minerals and fluids when the minerals are formed. Therefore, different stages of pyrite trace element information can be used to predict the type of deposit. In addition, the related research has a long history, and a large amount of data have been accumulated. However, due to the limitations of research ideas and methods, traditional methods often have problems of no solution or multiple solutions. With the development and promotion of big data ideas, researchers have tried to use machine learning algorithms to solve such problems, and have achieved good results. In this study, based on the idea of deep learning, a data set of trace elements of pyrite was established and subjected to in- depth research. By means of “component data visualization” and “data enhancement”, it solved the problems encountered by predecessors when using deep learning methods for such classification tasks. Previous attempts have pointed to the data imbalance problem and the inability of the convolutional neural network to directly read the data. This paper compares and analyzes the accuracy and effect of gold ore type classification task using pyrite trace element composition data based on four convolutional neural network models (Mobilenet V2, Resnet 50, VGG 16 and VIT). It is found that the convolutional neural network algorithm based on the pyrite composition data can perform more accurate classification tasks for different types of gold deposits. This method has higher accuracy and generalization ability than the traditional graphical method. By predicting the type of gold deposits, it can save costs for prospecting exploration and deep prediction work, and also provide deep learning for the application and promotion of geological and mineral resources research. The new idea has certain promotion value.