Applying Convolutional Neural Networks To Identify Lithofacies Of Large-N Cores From The Permian Basin And Gulf Of Mexico: The Importance Of The Quantity And Quality Of Training Data

Convolutional neural networks (CNNs), one of the most widely employed deep learning techniques, have achieved great success in image recognition. However, few attempts have been made in sedimentary studies, partially because it is challenging to generate a large-scale training database for sedimentary data. We compile similar to 32,000 images with interpreted facies from similar to 3200 ft (similar to 1000 m) of cores from the Permian Basin and Gulf of Mexico. This database is used to train and evaluate a CNN model predicting the facies from core images. The best learned model achieves 83% accuracy when evaluated by the independent testing data. More importantly, we analyze the impacts of sample sizes on the prediction accuracy to understand how many samples is needed for a model to return satisfied performance. Accuracy does increase as sample number increases but even the learned model trained by 1% of available data (n = 300) can return 70% accuracy. This result suggests the deep learning model is able to provide fast sedimentary analysis and accelerate the core description processes with small amounts of training dataset. We also show that the model trained from the Permian Basin data set fails to predict the facies of the Gulf of Mexico cores because the two data sets are in different depositional environments. Therefore, a high-quality training database covering different depositional environments is critical in applying artificial intelligence in facies detection. This study suggests the model trained by small amounts of high quality data can aid human interpretation. It can efficiently provide basic information such as bed thickness, lithology, and net-to-gross ratio to free up geologists to conduct more complex tasks such as interpreting depositional environments.