A Probabilistic Machine Learning based Framework for Lithology Prediction on Well Log Data
Author: Andrean Satria, Mordekhai*, Taufan Rusady
mordekhai.mordekhai@halliburton.com
Lithology prediction is one of the most important processes in petrophysical workflow since it is useful for knowing the prospective reservoir zone in the target well. Unfortunately, this process sometimes takes a long time and results in inaccurate interpretations due to the well data that has various mnemonics, the massive amount of data, and the inconsistency in manual interpretation. We present an assisted lithology interpretation framework with additional feature to compute prior and posterior probabilities during lithology prediction to help geoscientists if there are some irrelevant prediction results. We use various references to oil and gas basins data around the world resulting more than 60 pre-built models included in this framework. The framework can select model automatically based on the similarity between the well log curve in test data and references data. Besides being able to provide reliable and accurate results, this framework is a cloud based and has a centralized database. These features can accelerate collaboration and integration between users in predicting lithology. The data used in this study is from one of the most productive oil and gas field that has a varied number of wells and lithology. Based on these characteristics, this field is considered suitable for providing an objective judgement. This application is proven to be able to provide reliable results by producing prediction accuracy and F1 score above 0.6 using an automated model. This framework can also assist geoscientists to interpret exploration wells that do not yet have a valid lithology label.
Keywords: Probabilistic, Machine Learning, Lithology Prediction
View Full ManuscriptDownload