ED Sciences et environnements
Deep Learning for Carbonate Rocks Petrography
by Axel RANSINANGUE (Environnements et Paléoenvironnements Océaniques et Continentaux)
The defense will take place at 14h00 - Amphithéâtre 2 Rue Thomas Edison, 64000 Pau
in front of the jury composed of
- Raphaël BOURILLOT - Professeur - Université de Bordeaux - Directeur de these
- Cedric JOHN - Professeur - Queen Mary University of London - Rapporteur
- Stephen LOKIER - Professeur associé - University of Derby - Rapporteur
- Yannick BERTHOUMIEU - Professeur - Université de Bordeaux - Examinateur
- FOUBERT ANNELEEN - Professeure - Université de Fribourg - Examinateur
- Emmanuelle VENNIN - Professeure - Université de Bourgogne - Examinateur
- Richard LABOURDETTE - Docteur - TotalEnergies - CoDirecteur de these
- Nesrine CHEHATA - Professeure - Université de Bordeaux - CoDirecteur de these
This project aims to investigate deep learning capacity to enhance the identification of factors that may impact the dynamic behavior of carbonate reservoirs. This is achieved through the integration of comprehensive datasets comprising images acquired over decades from hydrocarbon carbonate reservoirs and identified outcrop analogues. Convolutional neural networks represent natural candidates for analyzing the thousands of available thin sections. This thesis developed a suite of complementary methods for their quantitative and automated description. An optimal subdivision strategy for high-resolution images was established to enable multi-scale characterization of this data type, closely approximating traditional microscopic analysis methods. The approach combines representation learning through a rotation-invariant variational autoencoder with automatic clustering, achieving a two-fold acceleration in labeling time while forming geologically coherent groups based primarily on R. Dunham's classifications for limestones, R. Folk's classifications for dolomites, and the incorporation of secondary minerals such as anhydrite facies. The model can then be refined for these subsidiary classification tasks, enabling semantic description of the samples studied (>88% mAP). To address the heterogeneity of existing petrographic descriptions and the lack of precise quantification on real data, two methods were developed: SynSection procedurally generates training data by simulating depositional processes (+20% mAP) in grain limestones, while SynSection2 unifies synthetic-to-realistic transformation with segmentation (+5% additional mAP). These approaches jointly enable automatic segmentation of elements (bioclasts, intraclasts, ooids, and peloids) and the creation of a homogeneous quantified database. Deep semantic segmentation applied to granular limestones provides access to quantification of relative proportions and grain size distribution analysis. The developed metrics achieve determination coefficients (R²) of 0.78 for mean grain size and 0.72 for sorting. This integrated methodology enables standardized and reproducible description following established petrographic nomenclatures and can be superimposed on pore network analysis. Automated quantification of pore types reveals heterogeneities that facilitate comparative analysis between samples. This approach constitutes a powerful tool for quantitative petrographic analysis, capable of incorporating local refinements to fully exploit deep learning's potential in characterizing carbonate systems and understanding associated petrophysical responses.
ED Sociétés, Politique, Santé Publique
University teaching practices in light of disciplinary pedagogical cultures: A located approach to the staging of teaching at Bachelor's level
by Maëlle TOURNEUR (Laboratoire Cultures et Diffusion des Savoirs)
The defense will take place at 14h00 - Amphithéâtre Durkheim Campus Victoire, 3 pl. de la Victoire 33000 Bordeaux
in front of the jury composed of
- Christophe ROINE - Professeur des universités - Université de Bordeaux - Directeur de these
- Saeed PAIVANDI - Professeur des universités - Université de Lorraine - Rapporteur
- Joris THIEVENAZ - Professeur des universités - Université Paris-Est Créteil - Rapporteur
- Marthe-Aline JUTAND - Maîtresse de conférences - Université de Bordeaux - Examinateur
- Sabine KAHN - Professeure - Université Libre de Bruxelles - Examinateur
- Barbara PENTIMALLI - Professeure - Université du Québec à Montréal - Examinateur
- Bernard SARRAZY - Professeur émérite - Université de Bordeaux - Examinateur
This study focuses on teaching practices at university, in a context of institutionalization and development of training programs for teaching practices. How do teachers stage their classes? If the recommendations provided in these training programs are intended to optimize learning, can they be disseminated independently of disciplinary specificities and teaching content? Conducted across seven disciplines and three course formats at the third-year level, this research aims to analyze teaching practices through the study of teachers' discursive practices, their teaching materials, their physical and bodily presence, and their interactions with students. It then aims to identify the logic underlying these practices. Based on observations of class sessions and interviews with teachers and students, this work highlights the dynamic, located and co-constructed dimension of teaching practices that are embedded in communities of practice, but also within disciplinary teaching cultures. The aim is to examine the pedagogical and didactic challenges of staging university teaching in real-life situations, while questioning generic and prescriptive discourse in relation to actual practices.