Permeability logs are critical data used to assess reservoir quality, which can impede the confidence of reservoir productivity without proper characterization. In particular, this can be significant in shale-sand sequence reservoirs that have heterogeneous lithology. Fortunately, permeability logs can be derived from Stoneley waves that are extracted from conventional full-waveform sonic logs. Because sonic logs are not always available in old wells, exploring alternatives to predict permeability is essential. The artificial neural network method is one option that provides elements to develop a model to predict permeability logs in offset wells. Consequently, a model is constructed using standard logs as input data and a Stoneley-wave permeability log as the desired output. These logs are obtained from a new well containing a complete suite of logs including full-waveform sonic data and a permeability log. The open-source Python library Keras provides the environment to generate the model. The model is applied to two nearby wells in which permeability logs do not exist. The borehole separation from the new well to the far well is 2055 ft, and the separation distance to the near well is 1370 ft. The logging data were acquired from wells located in an oil reservoir at Waggoner Ranch in northeast Texas. The reservoir geology is a sand-shale sequence intercepted by thin limestone markers at several depths. Permeability logs are predicted and correlated with other logs and their corresponding lithology. Sands partially saturated with hydrocarbons are identified by their permeability properties at the wells. The sands were previously characterized and mapped from impedance data using a nonlinear inversion of reflection seismic image constrained with well logs. The data analysis demonstrates the usefulness of the neural network approach to predict permeability in reservoirs where offset wells are available and only one well contains a permeability log.
Deep learning for predicting permeability logs in offset wells using an artificial neural network at a Waggoner Ranch reservoir, Texas
Jorge O. Parra
https://doi.org/10.1190/tle41030184.1