Leveraging Unsupervised Machine Learning on Seismic Images of Miocene Deposits in the Carpathian Foredeep Basin for Facies Classification

Abstract

The application of machine learning (ML) in seismic interpretation has become widely shared in geosciences, and this is due to its ability to be accurate, robust, and inexpensive. Using seismic attribute variety, the geological feature of subsurface data could be recognized in detail, where data can be presented in various physical properties depending on the purposes. However, manual interpretation could be time-consuming and take much effort. Therefore, machine learning is a robust method in determining facies on seismic data utilizing computer intelligence capabilities with minimal human intervention and a short time. Unsupervised learning when used can identify a pattern without a learning process. In other words, the data will separate itself automatically based on the method used without any labeling. In this research, we used the Gaussian Mixture Model (GMM) and K-Means Clustering methods to classify the distribution of facies in seismic images. The Gaussian Mixture model is defined as a model-based probabilistic formulation utilizing the number and shape of clusters determined in a more objective way using a Bayesian framework taking into account the possibility and complexity of the model. In contrast, K-Means Clustering is an iterative algorithm that is capable to group data into features or variables which has a similar characteristic, with K as a cluster having the nearest means. The primary motivation of this research is to increase our understanding of unsupervised machine learning, including the Gaussian Mixture Model (GMM) and the K-Means Clustering method for facies classification. These two methods were applied to selected seismic attributes with a target zone in the Miocene gas-rich formation, Carpathian Foredeep Basin, Poland. All these attributes have different parameters that might be correlated, and thus, to increase their statistical values, we applied Principal Component Analysis (PCA) as the dimensionality reduction tool. The key takeaways are on the distribution of sandstone, mudstone, and shale facies, where we identified that the Gaussian Mixture Model provides more convincing results than K-Means Clustering in the sandstone formation containing gas, which can be indicated as one group unit of a cluster.