_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper IPTC 22867, “Automatic Lithology Classification of Cuttings With Deep Learning,” by Takashi Nanjo, Akira Ebitani, and Kazuaki Ishikawa, Japan Organization for Metals and Energy Security, et al. The paper has not been peer reviewed. Copyright 2023 International Petroleum Technology Conference. Reproduced by permission. _ Wellsite geologists spend approximately 70% of their time on cuttings descriptions. In addition, two or three wellsite geologists generally are assigned to a drilling campaign, to be replaced at the end of a shift. Machine-learning (ML) and artificial-intelligence (AI) techniques have the potential to solve these issues because of their advantages in prediction speed, objectivity, and consistency. The authors’ aim is to automate the task of cuttings descriptions with these techniques. A trained model for cuttings description has the potential to realize quantitative, high-speed cuttings description. Methods and Materials Six wells were selected for this study (Poseidon-1, Ichthys2A ST-1, Ichthys2A ST-2, Dinichthys north1, Ichthys north1, and Ichthys north1 ST1). The wells were drilled in the Browse Basin through various lithologies. The authors focused on four lithologies (carbonate, sandstone, mudstone, and volcanic) and collected the cuttings from these lithologies. The total number of cuttings was 160 (21 carbonate, 58 sandstone, 52 mudstone, and 29 volcanic). The cuttings were used in their original dry condition. Each cutting was transported to a Petri plate in a random and sparse layout using tweezers. A stereomicroscope and a digital camera were prepared. Magnification was 6.3X, and picture size was 1,600 pixels in width and 1,200 pixels in height. Pictures were taken randomly, with approximately 20 photographs for each sample. The annotation was conducted by geologists using annotation software. Five label classes were used (sandstone, mudstone, carbonate, volcanic, and background). The total number of annotation data was 1,978. The annotated data were split as training label data, validation label data, and test label data. The pictures were also split as training data, validation data, and test data. The picture and label data were paired. In this study, 1,496 pictures and label data were used as training data sets, 320 pictures and label data were used as validation data, and 162 pictures and label data were used as test data. Four architectures of semantic segmentation were created (PSPNet, Unet, FPN, and Linknet), including two different networks (ResNet152 and EfficientNetB7), and training models were created. The total number of trained models was eight. As a result, the combination of PSPNet and EfficientNetB7 showed the best performance. The combination of PSPNet and EfficientNetB7 was chosen as the architecture of this project. Hyperparameters were chosen in the best practice case. Mean intersection over union (IOU) was used as the model evaluation index.
Read full abstract