Capturing interpretational uncertainty of depositional environments with Artificial Intelligence

Abstract

Geological interpretations are always linked with interpretational and conceptual uncertainty, which is difficult to elicit and quantify, often creating unquantified risks for understanding the subsurface. The complexity and variability of geological systems may lead geologists to analyse the same data and arrive at different conclusions based on their subjective interpretations, personal expertise, or biases. In order to address the associated uncertainty, it is valuable to consider multiple plausible interpretations of outcrop data and acknowledge the degree of ambiguity associated with each interpretation. By examining a diverse range of outcrop analogues, it becomes possible to derive multiple potential geological interpretations and identify variations within and across depositional systems. This thesis proposes a new AI system that learns valuable geological information from surface data (outcrop images), transfers this knowledge to the fragmented data of the subsurface (core data), and finally, links all the extracted information with the geological literature to produce plausible interpretations of the depositional environment based on a single outcrop image. To identify patterns and geological features within image data, three Supervised Learning Computer Vision techniques were employed: Image Classification, Object Detection, and Instance Segmentation. Natural Language Processing was utilised to extract geological features from textual information from heritage geological texts, thus complementing the analysis. Lastly, a custom Neural Network was deployed to assimilate the gathered information into meaningful sequences, apply geological constraints to these sequences, and generate multiple plausible interpretational scenarios, ranked in descending order of probability. The results of this study demonstrate that combining approaches from different areas of Artificial Intelligence within cross-disciplinary workflows under the umbrella of a broader AI system holds significant potential for subsurface characterization, better risk analysis, and potentially enhancing decision-making under uncertain conditions during subsurface exploration stages.