The Fossil Frontier: An answer to the 3-billion fossil question

Microfossil analysis is important in subsurface mapping, for example to match strata between wells. This analysis is currently conducted by specialist geoscientists who manually investigate large numbers of physical samples with the aim of identifying informative microfossil species and genera. The current digitalization of large volumes of microfossil samples that is being conducted by the Norwegian Offshore Directorate, paired with AI development, opens up new opportunities for automating parts of the analysis to help the geologist in the analysis. Unsupervised representation learning is a research area in Artificial Intelligence (AI) that lies at the core of this challenge, as this way of learning can create useful image representations by utilizing large volumes of data without requiring labels. Previous work has presented good results for the classification of a limited number of classes, but there are still challenges related to classification in realistic settings where additional unknown species are present. In this paper, we connect unsupervised representation learning and uncertainty estimation and create a comprehensive tool to automate microfossil analysis. We present our methodology and results in three parts. In the first part, we train several AI models from scratch using state-of-the-art self-supervised learning methods, obtaining excellent results compared against state-of-the-art foundation models for image classification and content-based image retrieval. In the second part, we develop a method based on conformal prediction which enables our classifier to handle a large pool of images containing both in-distribution and out-of-distribution data, while at the same time allowing us to create error estimates to control the uncertainty of the prediction sets. In the third part, we use our method to create distribution charts of fossils for a range of genera in multiple wells. • Train a microfossil model with superior classification and CBIR performance. • Develop novelty detection using conformal prediction for reliable uncertainty scores. • Create genus count charts for high-level well content analysis.