Effectiveness of Neural Kriging for Three-Dimensional Modeling of Sparse and Strongly Biased Distribution of Geological Data with Application to Seafloor Hydrothermal Mineralization

Three-dimensional modeling of geoscientific data of limited amounts and strongly biased locations is difficult and impractical using almost any method. To obtain a correct spatial model from data under such constraints, this study systematically demonstrated the effectiveness of neural kriging (NK), which is based on a deep neural network (DNN) with a semivariogram learning criterion. As a novel case study of resource geology, NK was applied to clarify the three-dimensional deposit structure of an active seafloor hydrothermal vent area, Izena Hole in the middle Okinawa Trough, using Cu, Zn, Pb, Ag, and Ba content data, geological columns, and resistivity data from drill core samples obtained at six drill sites aligned nearly E-W and two drill sites far from these six. Two high-content zones clearly appeared for all five elements, in the sulfide mound and underlying sediment, and about 30 m below the seafloor as a stratiform shape. Advantages of NK were demonstrated as follows. The NK results showed the locations of these high-content zones and their horizontal extent along the pathway of hydrothermal fluids more clearly than the result of typical geostatistical simulation, turning bands simulation (TBS), by enabling content estimations even in areas without data. NK also estimated the uppermost high-content zones more accurately than the ordinal DNN methods or TBS. Next, NK was tested for a region 20 times wider than the first target, and the resultant Cu and Zn content models showed indistinct horizontal linear features and vertical continuity, which may have been caused by the fault structure and small amounts of metal precipitation in the deep part, respectively. Thus, one advantage of NK is that its spatial model can contribute to constructing hypotheses of transport and enrichment mechanisms of target metals.

Automated summary

This study applied neural kriging (NK), a deep neural network (DNN) based on a semivariogram learning criterion, to model the three-dimensional deposit structure of an active seafloor hydrothermal vent area. The results showed that NK outperformed traditional geostatistical simulation methods in predicting the location and extent of high-content zones. The spatial model generated by NK can contribute to the understanding of transport and enrichment mechanisms of target metals.