چكيده به لاتين
Gas reservoir development and estimation of rock properties are highly dependent on lithological classification, which can be difficult, time-consuming and error-prone. In this study, a new deep learning-based approach is developed for fast, accurate and efficient lithology prediction in a gas field from standard well data. These logs are converted into 2D images using two proposed approaches, shallow images (SI) and deep images (DIs), where the pixels in these images effectively represent the relationships between different logs. For this purpose, we developed residual convolutional neural networks (ResCNN) named SIs-ResCNN 2D and DIs-ResCNN 2D. The input data for DIs-ResCNN 2D are initially vectorized images where the order of the reports is somehow repeated, ensuring that each binary combination occurs only once. This led to the integration of correlations between logs in the pixels of the generated images, along with the integration of unique binary combinations of logs. Based on the evaluation criteria, DIs-ResCNN 2D performed better than other proposed methods. The balanced DIs-ResCNN 2D model achieved an accuracy of 93% and an F1 score of 80% in the test well, which highlights the importance of data balance during CNN model training.
The thermal properties of sedimentary rocks, including thermal conductivity, thermal diffusivity, and specific heat capacity, are essential for predicting the thermal behavior of geosystems for a wide range of geoenergy applications, as well as for fundamental research. In this study, we develop a new open access tool for calculating continuous thermal profiles from data from standard geophysical logs. This tool is based on a physical modeling approach to extract basic synthetic datasets from mineralogy for the main sedimentary rock groups (clastic rocks, carbonates and evaporites) and use different machine learning methods to extract relationships between thermal properties. Each combination of standard log data (sound, neutron, density and shale volume fraction) is used as input. For different log combinations, we trained four different regression models (Linear, AdaBoost, Random Forest and XGBoost) on 80% of the data (chosen randomly). The remaining 20% of the data was used for statistical validation. For each main sedimentary group, we checked the significance of the characteristic of each input variable. In order to confirm the predictive power of the model, real data obtained from laboratory measurements performed on cores were used.
