Making Sense of Multi-Attribute Machine Learning: Gaining Seismic Interpretation Insights with SHAP for Unsupervised Clustering

Abstract

In seismic multi-attribute analysis, unsupervised machine learning helps identify patterns inherent in the data across the multiple attribute dimensions. It summarizes the information intuitively for appraisal of the underlying geological objects responsible for the different seismic signal expressions. Conventionally, though, unsupervised machine learning products—in the form of volumes and maps—serves mainly as leads for the seismic interpreter, while applying further seismic attribute theory for interpretation still requires going back to the original input attribute volumes. This points to a deeper challenge: unsupervised methods inherently lack an interpretability layer, leaving the model's internal reasoning opaque to the interpreter.Explainability frameworks such as SHapley Additive exPlanations (SHAP) have gained traction in supervised learning, where a well-defined output metric—typically a prediction probability or classification score—enables the attribution of input feature contributions. Without an equivalent metric, unsupervised models have largely been outside the scope of SHAP-based interpretability. To bridge this gap, this study utilizes probability distributions from a Gaussian Mixture Model (GMM) as a proxy metric for cluster membership confidence, effectively applying SHAP to an unsupervised multi-attribute machine learning workflow. This integration enables analysis into how individual seismic attributes contribute to each cluster assignment. Which means each appraised cluster is then accompanied by geophysically meaningful information that highlights the key signal expressions and attributes characterizing it. This added interpretability works to equip the geoscientist with data-driven quantitative means to support their interpretation, while simultaneously grounding the machine learning results back to seismic attribute theory. In this study, we found that the SHAP graphical descriptions align with seismic attribute characterization of geological targets in the Taranaki Basin, showing how the cluster is expressed in each attribute space. The graphical descriptions enhance the role of the unsupervised machine learning workflows in seismic multi-attribute analysis and provide intuition-driven mechanisms for quality control and trust in machine learning products.