3D Mineral Prospectivity Modeling Based On Machine Learning

The traditional 2D mineral prospectivity often requires a large number of known ore deposits,yet is unable to provide deeper region information,which restricts its effectiveness in the covered areas and the areas with insufficient known ore deposits.With the increase in resource demand and exploration difficulty,more single large deposits need to carry out 3D mineral prospectivity in the depth and edges,and the increasingly matured 3D geological modeling and 3D geophysical inversion technology can provid 3D mineral prospectivity with multiple data requiremnts.Also,the extraordinary data mining ability of machine learning can provide strong technical support for 3D mineral prospectivity.Therefore,it is urgent to carry out 3D mineral prospectivity research and case studiesbased on machine learning with the help of various spatial analysis methods.The Zhuxi tungsten deposit is the largest tungsten deposit discovered in the world.As a strategic resource,it is necessary to search for more tungsten seposites and ensure their security.In order to achieve this goal,results of 3D mineral prospectivity of its depth,edges and periphery are required.Based on the completed 3D geological modeling on six maps and 3D inversion of gravity,magnetic,electrical and seismic data in this area,The remaining five groups of characteristics,including residual density,magnetic susceptibility,resistivity,P-wave velocity and lithology,are divided into data set 1 including all five groups and data set 2 including only four groups of pure physical attributes by using the same mesh generation.Then,the feature attributes of the corresponding samples are extracted according to the spatial position of the known ore body or not,and the known samples are divided into the training set(75%)and the test set(25%).The training samples are trained by four machine learning algorithms: k-nearest neighbor(KNN),(error)back propagation neural network(BPNN),support vector machine(SVM)and random forest(RF),and the mean square error is calculated by grid search method and 10-fold cross validation to determine the best combination of parameters,and then carried out the classification and regression prediction of the 3D data of the whole region,and obtained a number of 3D mineral prospectivity models.The accuracy of each model is calculated by confusion matrix in classification,and the advantages and disadvantages of the training model are reflected by the receiver operation characteristic(ROC)curve in regression.Since ROC curve is not the only standard to judge the model,in order to avoid overfitting of training samples,the performance of the objective response model in prediction is evaluated by BPNN,SVM and RF model after the statistics of each model stage value The results show that the performance of the models on dataset 1 and dataset 2 is RF,SVM,and BPNN model.In order to better improve the prediction performance of the model,the four algorithms mentioned are combined to improve the prediction results by reducing the deviation of a single algorithm.Considering the different performance of different algorithms for different actual models and actual data,a weighted fusion method is proposed.In classification,the weight of each algorithm is determined based on the ratio between the accuracy of the test set and the proportion of the global prediction of 1;in regression,the weight of each algorithm is determined based on the capture efficiency of the global prediction.This algorithm overcomes the overfitting of the model to a certain extent when training known samples,and is more stable than the fusion model based on equal weights.Then,the three groups of fusion models with the same shape are combined into smoother and cleaner model for evaluation and interpretation of the prospective area.Based on this,six first level prospective areas are planned,which are the Zhuxi(T1),Henglu(T2),Taqian(T3),Lingang(T4),Yongshan(T5)and Zhenzhushan(T6).According to the prediction results,the vertical structural slices,the inevitable relationship between mineralization and nappe structure is disscussed,and the slices of different depths reflect the vertical changes of metallogenic prospect area.The projection position of the 3D prediction results and the previous 2D prediction results on the surface is relatively consistent,which proves the feasibility of this mineral prospectivity.At the same time,two new prospective areas of the 3D prediction may be the next prospecting direction.The relationship between the first level prospective area and the surface faults highlights the ore controlling effect of the faults,and the contact relationship between the first level prospective area and the granite shows that the oreforming heat source comes from the deep granite.The analysis of the physical properties of the first level prospective area objectively reflects that the mineral prospectivity also conforms to the theoretical basis of similarity analogy,and the gravity and magnetic response of the surface highlights that the anomaly cascade zone should be the focus of attention.The results show that on the basis of multi geophysical 3D inversion and 3D geological modeling,this method of 3D mineral prospectivity is expected to solve the obstacle of current mineral prospectivity,the difficulty in advancing to 3D,with the help of machine learning as well as to greatly improve the exploration efficiency and reduce the drilling risk.