Why Machine Learning Lithofacies Prediction will Transform Reservoir Characterization

Abstract This presentation introduces a modern machine learning (ML) workflow for predicting lithofacies that provides oil and gas producers with an efficient and cost-effective alternative to traditional seismic inversion methods. Using Self-Organized Maps (SOM), an unsupervised ML technique, this approach clusters seismic attributes to perform a detailed stratigraphic analysis of the subsurface. The resulting stratigraphic interpretation is then integrated with lithofacies information from petrophysical logs, combining independent data sources to improve prediction accuracy. Unlike conventional inversion techniques, this methodology does not rely on deterministic physical models, nor does it require time-consuming parameterization such as wavelet extraction or low-frequency trend determination. Identifying natural patterns in seismic data at the seismic sample interval can resolve both thin and thick beds with greater flexibility. The neurons generated by the SOM process share interpretive qualities with inverted data. However, ML-predicted lithofacies show sharper boundaries, both vertically and horizontally, due to the single-sample resolution of SOM. Demonstrations in the Paradise® software, supported by real-world case studies like the Niobrara formation, show that this ML-driven workflow accelerates subsurface modeling, reduces costs, and improves reservoir understanding. Producers benefit from more precise identification of lithology and sparse porous sands, enabling better well placement and more effective reservoir development. By leveraging modern computing power and advanced visualization, this workflow empowers geoscientists to describe the subsurface more comprehensively and make faster, data-driven decisions that directly impact field performance. #machinelearning #ai #lithofacies #geoscience #webinar