A review of machine learning applications to geophysical logging inversion of unconventional gas reservoir parameters

Abstract

Reservoir parameters are crucial indicators for reservoir evaluation and development and provide insights into long-term reservoir behavior. The primary methods for evaluating these parameters include direct core observations, experimental testing, and indirect evaluation techniques. Since its introduction, geophysical logging has been used to evaluate and invert reservoir parameters owing to its wide coverage. With an increasing focus on unconventional natural gas reservoirs, more refined reservoir evaluations and multiparameter analyses are required for their development to address the complex and microscopic models differing from those of the conventional petroleum reservoirs. Geophysical logging is important in several unconventional fields. Machine learning (ML) was used in unconventional gas reservoirs as an effective method to establish relationships between parameters and logging features. However, the accuracy of evaluating storage layers using a single ML method is limited. Studies focusing only on algorithm updates and indicator values are problematic in terms of interpretability and production applications. A need to standardize the use of algorithms and introduce validation comparisons such as geological methods is evident. In this study, we reviewed ML algorithms and models commonly used for logging inversion applications. The current research status and issues were analyzed for different unconventional gas reservoir parameters. Our findings emphasize the importance of combining geological and other methods for logging inversion using ML. We also used the random forest algorithm to accurately predict the reservoir porosity, gas content, coal structure, and macrolithotypes. Combined with established permeability and vitrinite reflectance models, factor analysis was used to comprehensively analyze and evaluate the coalbed methane reservoirs in the study area. In our assessment of the challenges and future work on ML-based inversion, we observed a clear advantage for ML algorithms under geologically validated methods and experimental control. ML has great potential for optimizing the application of logging inversion for unconventional reservoir parameters.