Learning meaningful geological fields representations by self-supervised methods

Grigoriy M. Shutov, Klim Goldshtein, Marsel Khayrullin, Iuliia Uzhegova, Konoshonkina Natalia Yuryevna, Stanislav Slivkin, Gleb Shishaev, Boris Belozerov, Evgeny Burnaev, Vladimir Vanovskiy

Keywords: self-supervised learning, geological modeling, geological representations, learning representations

TL;DR: Learning meaningful self-supervised representations of geological fields helps to train better generative models for downstream tasks

Abstract: Geological fields are challenging to model as they reflect complex multi-scale sedimentation processes during the whole history of Earth’s existence. The geological objects that can be found in such fields have very different nature and scale and are often represented by statistical means. In this work we investigate the possibility to obtain meaningful representations for geological fields of different kinds with the help of self-supervised learning (SSL). Specifically, we adapt a self-supervised vision transformer architecture DINOv2 to a dataset of more than twenty thousands synthetic 3D geological cubes describing four sedimentary environments to learn their representations without labels. We show that learned embeddings cluster by sedimentary environment type and have high correlation with important geostatistical properties, such as variograms. When used as features for classification and regression, they match or exceed the performance of supervised CNNs trained from scratch. Our results demonstrate that SSL can capture meaningful geological structure in 3D geological data and serve as a strong foundation for downstream tasks, reducing the need for expensive labeled datasets in geological modeling.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2026/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 25226