Machine Learning Prediction of Formation Evaluation Logs in the Gulf of Mexico

The objective of the paper is to describe the application of artificial intelligence software to predict formation evaluation logs (compressional sonic, shear sonic and density) using only gamma ray, and resistivity log data and drilling dynamics data as received by the electronic drilling recorder (EDR). The software was applied real-time as a well was being drilled in deepwater Gulf of Mexico. Thorough examination and conditioning of EDR and wireline data give way to a training model construction for the artificial neural network (ANN) using full suites of log-data in offset wells. Next, a neural network architecture and associated hyperparameters are chosen and tested. The fully trained and validated model is applied to the gamma ray, resistivity and EDR of the target well while drilling. Real-time EDR and wireline data flow via WITSML from rig to cloud and data is delivered to the client. The results of the study indicate the simulated log data were comparable to those measured from conventional logging tools over the study area. In both blind well tests the density agreed with the conventional log results within 1.1 % and the compressional within 2.51 % (Figure 1). Each of these is well within the range of variance expected of repeat runs of a conventional logging tool. A primary driver for near real-time logs was to confirm structural depth of the target sands along the well bore. There was a depleted sand below the expected TD of the well that, if encountered, could have led to total losses and possible loss of the wellbore. It was critical to have real-time logs to characterize the sands above the depleted sand, using every possible petrophysical and geologic character to refine the log correlation. This integration of all the logs provided the best interpretation of the sand quality and led toward the completion decision. AI-based logs are a highly cost-effective alternative to LWD logging. It presents an environmentally friendly approach as there is no logging personnel on-site and no expensive and potentially dangerous nuclear sources in the hole The deployment of this patented, machine learning-driven, real-time simulation of formation evaluation logs is unique in using only gamma ray, resistivity and drilling data. It is particularly useful in the overburden section where formation evaluation tools are often not run for cost reasons, in side-tracks, in HP/HT settings and operational risk mitigation. It provides additive data for other petrophysical/QI/rock property analyses including seismic inversion, shale content, porosity, log QC/editing, real-time LWD, drilling optimization, etc.