Application of Class-Based Machine Learning for Potential Hydrocarbon Zones Identification: A Case Study

Abstract

At times, petrophysicists are expected to evaluate potential of the well in time-constraint situations while maintaining consistency of the parameters and interpretation. Other than that, some challenges may also occur when working with older wells where the dataset are not as complete as current wells and processing parameters are not transferable. In this case study, class-based machine learning (CBML) approach is used to perform petrophysical evaluation to identify potential hydrocarbon zones in the target wells. The objective is to find solution to improve efficiency and consistency in those challenging situations. A class-based machine learning (CBML) workflow uses cross-entropy clustering (CEC)-Gaussian mixture model (GMM)- hidden Markov model (HMM) workflow that identifies locally stationary zones sharing similar statistical properties in logs, and then propagates zonation information from training wells to other wells (Jain, et al., 2019). The workflow is divided into two (2) main steps: training and prediction. Key wells which best represent the formation in the field are used to train the model. This approach automatically generates the number of cluster (class) using unsupervised or supervised depending on the input data. The model from key wells data is then used to reconstruct inputs and outputs along with uncertainty and outlier flags. This allows expert to QC and validate the generated class which is the most crucial part of the workflow. Once the model from the key wells has been built, it is applied to predict the same set of zones in the new wells that require interpretation and predict output curves. The result matched well over the good data interval with the petrophysical interpretation result from conventional approach. While in the bad interval, some discrepancies can be observed. The discrepancy was identified easily from the uncertainty and outlier flags which helps petrophysicists to identify which interval to fix or re-evaluate. Some requirements to condition the input were observed (no missing value over the input and outlier) to get the best result. A number of inputs used in the model need to be consistent over the set of wells used in the training and prediction target. This machine learning workflow speeds-up the petrophysical analysis process, reduce analyst bias and improve consistency result between one well to another within the same field. This machine learning application can also generate auto log QC, zonation class for rock typing also reconstructed logs which enrich the petrophysical interpretation even for wells with limited logs availability. This paper offers practical examples and lessons learned of CBML approach application to perform petrophysical evaluation and identify potential zones while being in time-constrained and limited resource situations.