A Comparative Study of Machine Learning Techniques for Facies Characterization
by Subhadeep Sarkar, Schlumberger
was presented on Wednesday the 6 th of February, 2019.
In petrophysical modeling, determination of lithofacies is a crucial part of the workflow as it helps to classify similar types of reservoir rocks in the spatial domain and allows identification of potential pay zones. Conventional methodologies of facies classification rely mostly on the visualization process of the human interpreters and is extremely tedious and prone to errors. With the onset of the machine learning era, this procedure can be significantly improved by the implementation of several supervised and unsupervised learning techniques. Three commonly used supervised machine learning algorithms, namely the Decision Trees, Random Forest (RF) and the Support Vector Machine (SVM) are used on a dataset associated with a complex lithology. The robustness of the three algorithms for different training conditions like small training sets or data unseen by the model are then compared. In terms of Facies classification, SVM is efficient in clustering data with small training sets than the Decision trees. On comparing the predictive error for the three different algorithms, one common observation was that over-modeling can affect the precision. Also, as expected, the models were more efficient when the predictive classes were fewer. However, a key to the facies classification task using supervised machine learning algorithm is to properly condition the training dataset, so that all the facies are well represented in the training phase. Also, to optimize the bias and variance, which is an inherent problem with most machine learning techniques, cross-validation needs to be performed to fine tune the algorithms.
Subhadeep Sarkar is a Petrophysicist at Schlumberger Software Integrated Solutions in Tananger. He received a B.Sc. in Physics from Presidency College, Kolkata in 2008 an M. Tech in Geophysics from Banaras Hindu University in 2011. At Schlumberger, he focuses on NMR, petrophysics, and acoustics data interpretation for a wide range of wireline and logging-while-drilling measurements. He has worked on various conventional and unconventional reservoirs in India, North America and Gulf of Mexico. His interests also include implementation of machine learning techniques in novel petrophysical workflows and their parameterization.