Abstract
Identifying lithofacies is fundamental to petroleum geology and engineering, offering crucial insights into the distribution, quality, and behavior of hydrocarbon reservoirs. These insights are vital for efficient exploration, development, and production strategies while minimizing operational risks. However, utilizing seismic data to identify lithofacies presents challenges due to limitations in resolution, interpretation complexity, and the indirect nature of seismic measurements. Seismic data primarily capture the elastic properties of subsurface materials rather than direct lithologic characteristics. While advanced techniques like amplitude variation with offset analysis and inversion, supported by robust rock physics models, have mitigated some of these challenges, they are often constrained by their low dimensionality, low resolution, and reliance on fitting physical models. We propose a novel data-driven, machine-learning methodology for predicting lithofacies from seismic and well data. Unlike deterministic approaches, our method does not require seismic and well data to conform to physical models. It reduces ambiguity and nonuniqueness by incorporating multiple attributes into the analysis and enhances interpretability by categorizing each seismic sample with the most probable lithofacies and their associated probability. We create a litho-stratigraphic model of the subsurface by integrating two independent data sources: natural clusters of seismic attributes and lithofacies from well-log data. This model delineates lithofacies at a granular level, both vertically and horizontally, within each seismic sample. The methodology has been applied in several geologic settings and has proven efficient and effective.
Introduction
Methodology
- Two unsupervised ML clustering algorithms were employed: self-organizing maps (SOMs) and k-means to enhance the analysis of seismic and petrophysical data. SOMs are particularly effective for high-dimensional data visualization and pattern recognition. By mapping multidimensional seismic attributes onto a two dimensional grid, SOMs preserve topological relationships, allowing the identification of natural clusters within the seismic data that correspond to different lithofacies or rock types.
- Real-World Application: The case study demonstrating the methodology was conducted over 100 square miles of recently acquired, high-quality, multiclient seismic data in the DJ Basin in Colorado, USA. The reservoir comprises 300 to 400 ft of thin chalk benches interlayered with high total organic carbon marly shale in the Smoky Hill Member of the late Cretaceous Niobrara Formation.
Problem Statement
-
Manual lithofacies interpretation is slow, subjective, and limited in scale.
-
Thin beds and subtle lithologic variations are often missed by traditional amplitude-based interpretation.
-
Increasing volume of seismic data and attributes makes it impractical for interpreters to evaluate comprehensively.
-
There is a need for a faster, unbiased, and more accurate workflow to predict lithofacies directly from seismic data.
Course of Action
-
Integrate multiattribute seismic analysis with machine learning to automate lithofacies prediction.
-
Apply principal component analysis (PCA) to focus on the most impactful attributes.
-
Deploy the workflow across 3D seismic data to produce continuous lithofacies predictions.
-
Validate and refine results by comparing with independent wells and geologic models.
