Automatic Lithology Classification of Cuttings with Deep Learning

Describing cuttings is routine work for wellsite geologists on a drill rig. The time-consuming nature of this analysis and the lack of consistency of the results between different interpreters are the two major concerns for this task. Wellsite geologists spend approximately 70% of their time on cuttings descriptions. In addition, 2 to 3 wellsite geologists are generally assigned to a drilling campaign, and they are replaced at the end of a shift. ML/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 ML/AI techniques. We are targeting four lithologies, namely sandstone, mudstone, volcanic (volcanic rocks), and carbonate (carbonate rocks). The cuttings were collected from six wells in the Browse Basin (Australia). Of these four lithologies, a total of 1978 cuttings images were taken under a stereomicroscope. We chose a semantic segmentation technique using Convolutional Neural Network (CNN) algorithms to perform the image classification task. The images were labelled using the open-source annotation software. This annotated data were used for the network training. The labelled images were split into training, validation, and test sets. The accuracy of the trained model was evaluated using the intersection-over-union metric (IOU). The mean IOU of the final model on the validation dataset was 82.3%. Prediction results on the cuttings that are represented by single lithologies are qualitatively very accurate. On the other hand, the prediction for the non-typical lithology (e.g., siltstone, dark-colored volcanic rocks, mixed lithology samples) has room for improvement. The fragments with similar textures (e.g., dark colored volcanic and dark mudstone) are complex for the CNN to identify. The final goal of our project is not only the lithology identification but also the quantitative estimation of lithology abundances in the cuttings. Additional model improvements, such as hyperparameter optimization and significantly more training data, are required to accomplish this task successfully. The trained model for cuttings description has the potential to realize quantitative and high-speed cuttings description. Well-trained AI/ML models have the potential to assist well site geologists by automating the cuttings description process simplifying, speeding up and improving the consistency on the rig floor.