Image logs such as Microresistivity give the interpreter an indication of bedding patterns and depositional environments. This allows him, after correlation with an analysis of selected curve shapes (SP, GR) to interpret the grain size along the well. Both image analysis (from image logs) and data aggregation are areas where Deep Neural Networks have proven efficient, but their accuracy in the field of grain size prediction is still limited, hence not useful yet for an interpreter. Since relevant and complete log data is usually only available on small intervals, networks are overly trained on hand-picked well sections, while ultimately predicting a softmax-activated vector of all grain size "probabilities".
Working with data from deepwater wells off the Gulf of Mexico, we have been able to account for uncertainties and scattering in the training data to produce a set of single-tasked, multi-input, multi-output classification CNNs. Combined, these networks produced more accurate results than their multi-category counterparts, with accuracy over 87% on test wells. Using the computed metrics (recall, precision) of these single networks, a confidence curve of the overall algorithm is produced, highlighting high-confidence high-accuracy areas (85% of test wells), as well as areas where human interpretation is required.
This approach combines Machine Learning prediction and uncertainty analysis to reshape, using deterministic methods, the "black box" predictions of the Neural Network. This provides a methodology which is more accurate than conventional Machine Learning, while also being aware of its own inaccuracies and highlighting them. It uses complex network architectures and data ingestion mechanisms, designed to generate results that can be exploited by a petrophysicist in charge of log interpretation.