カッティングスサンプルにおける岩相判別AIモデルの構築

<p>Japan Organization for Metals and Energy Security conducted a joint research project with the Commonwealth Scientific and Industrial Research Organization in Australia in 2021-2023 to automate cuttings descriptions using artificial intelligence - machine learning techniques. The project target was to investigate the possibility of minimizing the time for cuttings descriptions and establishing quantitative descriptions.</p><p>A machine-learning model was constructed to distinguish four lithologies（sandstone, mudstone, carbonate, and volcanics）in addition to the background from dried cuttings. The samples were placed in Petri dishes for observation under a stereomicroscope, and images were captured using a camera connected to the microscope. The captured 989 images were labeled with the lithologies using the open-source LabelMe annotation software. The labeled and captured images were paired to create a dataset. Subsequently, the dataset was divided into training, validation, and testing sets to construct the machine-learning model.</p><p>Eight machine-learning models were created and using four architectures（Unet, PSPNet, FPN, Linknet）and two backbone networks（EfficientNetB7 and ResNet152）for semantic segmentation approach. The combination of PSPNet and EfficientNetB7 showed the highest Intersection over Union（IOU）values. The IOU for each lithological type on the unseen test data were 81.4 % for carbonate, 53.4 % for mudstone, 67.4 % for sandstone, and 84.5 % for volcanics. IOU were higher for carbonate and volcanics and lower for sandstone and mudstone, because the rocks have consistent appearances in color and texture, making them easier to predict. However, siltstone, which is difficult to distinguish from sandstone and/or mudstone, and rocks with similar outlook, such as dark-colored volcanics and dark-colored mudstone, are more challenging for prediction.</p><p>In conclusion, it was possible to discriminate the four lithologies with a certain accuracy. While, other challenges, such as how to automate other items in cuttings description and practical application process on the rig, are recognized to automate cuttings description.</p>