Automated Subsurface Interpretation: Deep Learning Systems for Well Correlation and Reservoir Characterization

This CIFRE thesis was carried out in partnership between LISN and SLB under the direction of Frédéric BOULANGER, Professor CentraleSupelec (LMF), and the co-supervision of Francesca BUGIOTTI, Associate Professor, CentraleSupelec (LISN) and Sylvain WLODARCZYK, Senior Data Scientist, SLB.

The defense will be held in English.

Defense committee

• Genoveva Vargas Solar, Researcher (HDR), Université Claude Bernard Lyon 1 (Rapporteur)
• Oscar Romero, Professor, Universitat Politécnica de Catalunya (Rapporteur)
• Bich-Liên Doan, Professeur, HDR, Université Paris-Saclay (Examinatrice)
• Patrick Marcel, Professeur, Université d’Orléans   (Examinateur)
• Stefania Dumbrava, Associate Professor, Institut Polytechnique de Paris (Examinatrice)
• Julien Romero, Associate Professor, Institut Polytechnique de Paris (Examinateur)

Abstract

Accurate subsurface characterization is essential for reservoir modeling, carbon capture and storage, geothermal energy, and sustainable resource management. Well correlation is the alignment of geological formation tops across boreholes. It is traditionally done by manual interpretation of well logs, which is slow, subjective, and hard to scale. This thesis develops Deep Stratigraphic Inference tools, from localized pattern matching to global, physics-informed sequence modeling.

The work progresses through GeoTS, a modular LSTMLong short-term memory-CNN library for gamma ray log interpretation with AdaptiveCycle training and GradCAM explainability; GeoTS+, which extends this to multi-log inputs (density, resistivity, sonic, neutron) with log-adaptive inference and local and global context encoding; and LithoFormer, a sequence-to-sequence transformer framework that is the main contribution. Diagnostic studies show that graph neural networks blur vertical boundaries and naïve transformer regression is unstable for depth picking motivating LithoFormer’s PatchTST backbone, rotary positional embeddings, and geology-informed multi-task learning. On three real datasets, LithoFormer cuts median boundary error by ~90% and removes stratigraphic order violations. The thesis closes with a Geological Foundation Model vision for unified imputation, property regression, and stratigraphic segmentation at basin scale.