Legacy IODP Cores Data in the Era of Big Data

Core data is pivotal for understanding our planet’s past, present, and future. Despite this richness, extracting meaningful insights from core description poses significant challenges due to the inherent complexity and variability of the data, the amount of existing material, and the subjectivity of the interpreter. IODP (and the preceding programs) offers a rich, well labelled source of core images that can be used in machine learning and deep learning.

Focusing largely (but not exclusively) on carbonate rocks, characterized by their heterogeneity at all observational scales, I will discuss how my research group and I have pioneered the application of deep-learning computer vision to geological core interpretation. This technology transcends the traditional, tedious manual interpretations of cores, offering a rapid, and often more accurate, alternative for delineating depositional environments and sequence stratigraphy. Convolutional neural networks (CNNs) form the backbone of our approach, enabling us to process core data with unprecedented efficiency. I will show that these sophisticated models, when correctly trained and fed with substantial datasets, serve as invaluable tools for geologists, outpacing conventional methods in speed without compromising on precision.

Our early work was centred on transfer learning, an AI approach that adapts pre-existing models to new data. I will show that this remains one of the best way to train classification algorithms for geological dataset. But we also worked on generative algorithms that fill gaps in our sampling of core imagery: for instance, we use Generative Adversarial Networks (GANs) to transform the resistivity images from formation micro scanners into representations mirroring actual core photographs, thus enhancing the interpretability for geologists irrespective of their background in downhole tools.

We tackle the often-limiting factor of dataset size in two ways. First, we recourse to generative AI to oversample our training set. Second, we also explore semi-supervised learning techniques.  I will demonstrate that we successfully train models on core deformation images from IODP with minimal labelled data, achieving accuracy on par with, if not exceeding, that of transfer learning models.

None of our achievements would have been possible without the recourse to IODP data. Hence, this presentation serves as a clear illustration of the value of legacy IODP data for future geoscientists.