[S-GL17] Geochronology and Isotope Geology

[SGL17-P03] An experimental trial of auto-handpicking mineral separation for K–Ar dating using deep-learning micromanipulator system

*Seiko Yamasaki¹, Ayumu Miyakawa¹, Takuya Itaki¹ (1.Geological Survey of Japan, Advanced Industrial Science and Technology)

Keywords:K–Ar dating, Sample preparation, deep-learning, AI classification

For reliable K–Ar dating of volcanic rocks, mineral separation is important to remove phenocrysts that may contain excess ⁴⁰Ar. In general, phenocrysts grains that cannot be removed by magnetic separator and/or heavy-liquid separation are finally removed by hand-picking under a microscope, however, hand-picking is a time-consuming process. In this study, we conducted an experimental trial of auto-handpicking separation using deep-learning micromanipulator system at Geological Survey of Japan which has successfully used for classifying microfossils. A rock sample from Ontake Volcano was crushed, sieved to 180-250 μm and the separated groundmass fraction with a small amount of plagioclase grains was used for the experiment. In classifying microfossils, the feature of shape is used as a criterion, but this sample need to classify plagioclase and groundmass in color. On making training data, the original software resulted in classification that emphasized the shape of grains, so the collected individual images of grains were once extracted and classified using deep-learning software that newly developed in GSJ. The classification model with the training data showed the ability to classify plagioclase and groundmass grains with 95% accuracy. Although there are still some problems in classifying when a single grain contains plagioclase and groundmass part, the effectiveness of the system to mineral separation was confirmed.