Abstract: One of the most important applications of geochemical whole-rock analysis is to discriminate the tectonic settings for magma formation and properties of magmatic source areas through geochemical characteristics of magmatic rocks. This approach allows discrimination of tectonic setting of a given suite of magmatic rocks (basalt, granite, etc.) using whole-rock geochemical data including major and trace elemental and isotopic compositions. As a new application of artificial intelligence technique in the field of geochemistry, machine learning discrimination algorithm has gradually become a research tool supplementary to the classical discrimination diagram approach. However, the algorithms classification accuracy is affected by two main factors: high-dimensional data feature screening and multiple unknown parameter determination. To this end, we propose here a coupling discrimination method involving improved genetic algorithm and optimized neural network (GA-NNDM), based on genetic algorithm (GA) and neural network discrimination method (NNDM). The proposed method uses the feedback links between feature selection, parameter determination and classification performance. It treats classification accuracy as a fitness function and seeks the best feature subset and unknown parameters through iterative evolution. As a result, data features are reduced, unknown parameters are optimized and classification performance is improved. In addition, according to the published geochemical data of basalt samples, vertical and horizontal comparison experiments are set up through K-fold cross-validation method to verify accuracy, stability and extensibility of GA-NNDM in the application of basalt tectonic setting discrimination. Simulation results show that GA-NNDM has an excellent discrimination effect and generalization ability, with the overall classification accuracy near 90%. We conclude that, as a whole, GA-NNDM can be applied widely in geochemistry.

Key words: basalt, tectonic setting discrimination, neural network, genetic algorithm, feature selection, parameter optimization