Well log data analytics: overview of applications to improve subsurface characterisation

Wireline log datasets complemented with core measurements and expert interpretation are vital for accurate reservoir characterisation. In many cases, effective use of this information for predicting rock properties requires application of advanced data analytics (DA) techniques. We developed non-linear prediction models by combining data- and knowledge-driven methods. These models were used for predicting total organic carbon and electro-facies from basic wireline logs. Four DA approaches were utilised: unsupervised, supervised, semi-supervised and expert rule based. The unsupervised approach implements ensemble clustering for detecting variations in sedimentary sequences of the subsurface. The supervised approach predicts rock properties from well logs by applying ensemble learning that requires core data measurements. The semi-supervised approach builds a decision tree for iterative clustering of well logs to locate a specific facies and uses criteria determined by a petrophysicist for making decisions at each tree node whether to continue or stop the partitioning. The expert rule based approach combines clustering techniques at individual wells with an expert’s methodology of interpreting facies to determine field-wide rock characterisation. Here we overview the developed models and their applications to log data from offshore and onshore Australian wells. We discuss the deep thinking–shallow learning versus shallow thinking–deep learning approaches in reservoir modelling and highlight the importance of close collaboration of data analysts with domain experts.