Analysis of Clustering Machine Learning Algorithms and Generative Artificial Intelligence Tool for Visualization and Interpretation of Seismic Data

Abstract Seismic attribute analysis is often confounded by the complexity of the subsurface. Clustering machine learning algorithms reduce complexity and bring the most salient features of the seismic datasets to the fore. By conducting a comparative study of clustering algorithms applied to 2D seismic data, we can determine a set of optimal methods for visualizing particular seismic attributes. This paper presents an analysis of clustering algorithms, and a generative artificial intelligence-based large language model (LLM) customized for interrogating seismic data. Based on data partitioning, density and distribution category, K-means, DBSCAN, and Gaussian Mixture Modeling (GMM) algorithms were used to analyze amplitudes in 2D stacked seismic data. To assess cluster quality, performance metrics such as the Silhouette index and Davies-Bouldin Index were determined. The analysis revealed that seismic data can be effectively divided into clusters based on the amplitudes and frequencies along each of the traces and times. These clusters allow interpretation geophysicists to readily identify potential geohazards such as shallow gas, faults, or other amplitude anomalies. It was observed that K-means successfully divided the data into clusters in a manner which was sensitive to the overlapping data and noise. GMM exhibited a probabilistic feature of partitioning the data effectively. DBSCAN and GMM provided well distributed and identifiable clustering and fitting of amplitudes. The choice of clustering algorithm depends on the data size, data complexity, and performance metrics. Clustering is often regarded as a foundational technique in various types of data analysis. It may allow interpretation geophysicists to readily discern prominent and physically relevant amplitudes, amongst other seismic attributes. Further work, detailed in section 2 of this paper, focuses on developing a generative AI-based model that can function as a custom-made domain-interacting tool. This serves as a basis for developing additional tools through fine-tuning and additional training of AI models. These models then can be trained to predict subsurface structures and help processing/interpretation geophysicists to review and understand seismic data from a different perspective.