Phd defense on 13-11-2025

ED Mathématiques et Informatique

• Incremental Explainable AI for Multi-modal Data, Application to the Recognition of Risk Situations

  by Meghna Parameswaran AYYAR (LaBRI - Laboratoire Bordelais de Recherche en Informatique)

  The defense will take place at 14h30 - Amphithéâtre (050) LaBRI, 351, cours de la Libération F-33405, Talence cedex

  in front of the jury composed of

  • Jenny BENOIS-PINEAU - Full professor - Université de Bordeaux - Directeur de these
  • Akka ZEMMARI - Full professor - Université de Bordeaux - CoDirecteur de these
  • Pascal DESBARATS - Full professor - Université de Bordeaux - Examinateur
  • Damian GARREAU - Full professor - Julius-Maximilians-Universität Würzburg - Examinateur
  • Hichem SAHBI - Chargé de recherche - LIP6, CNRS - Rapporteur
  • Guillaume GRAVIER - Directeur de recherche - IRISA, INRIA - Rapporteur
  • Hervé JEGOU - Senior Research Director - Meta - Examinateur

  Summary

  Artificial Intelligence (AI) has become a cornerstone of modern technology, driving advances in computer vision and natural language processing. Despite these successes, current AI methods remain limited by their reliance on large static datasets, limited adaptability to dynamic environments, and a lack of interpretability due to their "black-box" nature. Unlike humans, who learn continuously while retaining prior knowledge, neural networks struggle with incremental adaptation and often suffer from catastrophic forgetting. Furthermore, the opacity of deep models poses challenges for trust, accountability, and human oversight. These issues underscore the importance of Incremental Learning, which enables models to learn continuously from data streams, and for explainable AI (XAI), which improves transparency and interpretability. This thesis addresses these challenges through three main contributions. First, we introduce a family of fast methods for online streaming learning in the one-sample-at-a-time setting, extending the Move-to-Data (MTD) approach previously developed at LaBRI. We provide a systematic convergence analysis of MTD and propose several variants, including Entropy-based MTD (EMTD), which leverages prediction entropy to select informative samples, and MTDR/EMTDR, which employ a conditional retargeting mechanism to mitigate model drift. A key novelty of MTD is that it updates the classifier directly without full backpropagation and performs a conditional single-step gradient correction for retargeting. Evaluations on public benchmarks and a real-world multimodal dataset demonstrate that our variants achieve competitive or superior performance while being much faster than traditional gradient-based approaches. Second, we focus on explainability for Vision Transformers (ViT). We adapt the FEM method, originally designed for CNNs, to the transformer paradigm by introducing Rollout-FEM (RFEM). RFEM uses transformer self-attention and applies statistical filtering to retain only the most informative weights. As RFEM is class-agnostic, we further propose RFEM-Class, a class-specific extension incorporating attention gradients to improve interpretability with respect to a target class. Evaluations on datasets containing human gaze fixations show that our methods achieve high plausibility scores, closely matching human visual perception. Our methods outperform other transformer explainers, such as Attention Rollout, Grad-CAM, and AttnLRP, while also demonstrating consistent stability, correctness, and robustness under perturbation-based evaluations. Third, we integrate explainability into the training process. We propose the Explain-to-Train (ET) framework, which uses RFEM explanations to guide models toward the most relevant input regions, improving performance across video, sensor, and multimodal datasets. Building upon this, we unify streaming learning and explainable training to introduce the Explain-to-Improve Streaming Learning (ESL) framework. ESL incorporates RFEM explanations with EMTDR to enhance data selection and model updates in online streaming settings. Finally, we extend our methods to unsupervised streaming test-time adaptation (TTA) by leveraging pseudo-labels from model predictions and adapting MTD and EMTD for this setting. This extension provides a promising direction toward practical and realistic unsupervised online adaptation. In summary, this thesis contributes (i) a family of Move-to-Data methods for fast online streaming learning, (ii) XAI methods tailored for ViT, and (iii) frameworks integrating explanations into training and streaming learning. It also lays the groundwork for future research on unsupervised test-time adaptation, demonstrating the potential to extend these methods to practical and realistic scenarios.

• A Diffeomorphic Mapping Approach for Model Order Reduction in Aerodynamics

  by Jon LABATUT (IMB - Institut de Mathématiques de Bordeaux)

  The defense will take place at 14h00 - Salle Contensou centre ONERA châtillon 29 Av. de la Division Leclerc 92320 Châtillon

  in front of the jury composed of

  • Angelo IOLLO - Professeur - Université de Bordeaux - Directeur de these
  • Damiano LOMBARDI - Directeur de recherche - INRIA Rocquencourt - Rapporteur
  • Sanderse BENJAMIN - Directeur de recherche - Centrum voor Wiskunde en Informatica (CWI) - Rapporteur
  • Astrid DECOENE - Professeure - Université de Bordeaux - Examinateur
  • Emmanuel TRéLAT - Professeur - Sorbonne Université (Paris 6) - Examinateur
  • Tommaso TADDEI - Professeur associé - Sapienza University of Rome - CoDirecteur de these

  Summary

  Parametric fluid dynamics simulations present significant computational challenges in aerodynamics, where high-fidelity computational fluid dynamics (CFD) models involve millions of degrees of freedom. Reduced-order models (ROMs) address this computa- tional burden by constructing low-dimensional approximations of the full-order system. However, standard projection-based approaches exhibit poor performance for advection- dominated flows. A proposed solution is to rely on coordinate transformation mappings to enhance ROMs compression. This thesis addresses the issue of developing a general framework for diffeomorphic mappings to align aerodynamics structures for application in coordinate transformation based ROMs. To answer this problem, we first establish the mathematical foundations of the method. For this, the mapping of interest is defined as the minimizer of an objective function. This constitutes the registration problem. The objective function combines a data misfit term, based on point set alignment of coherent structures, with a regularization term derived from differential operators. The mapping is defined from a parametrization of a velocity field. This guarantees diffeomorphic transformations in bounded CFD domains under tangent boundary conditions. The minimization problem is solved using a finite element discretization, gradient-based optimization, and an Expectation-Maximization algorithm for automatic coherent structure labeling. The methodology is validated on three representative test cases of increasing complex- ity: (i) coalescing Gaussian mixtures, to illustrate the alignment of merging structures; (ii) transonic Euler flow around a NACA0012 airfoil, demonstrating multi-structure registration; and (iii) viscous RANS flow around the ONERA M6 wing, showcasing the alignment of complex 3D lambda shocks. Across these cases, the proposed approach enhances ROM accuracy, and demonstrates robustness to aerodynamic problems.

ED Sciences Chimiques

• In vitro systems for bone tissue engineering: Biochemical multi-functionalization with mimetic peptides

  by Melissa KOSOVARI (Institut de Chimie & de Biologie des Membranes & des Nano-objets)

  The defense will take place at 14h30 - Salle des thèses Bâtiment C, r-d-c, 16 Avenue Léon Duguit, 33600 Pessac

  in front of the jury composed of

  • Marie-Christine DURRIEU - Directrice de recherche - Université de Bordeaux - Directeur de these
  • Farida DJOUAD - Directrice de recherche - The Institute for Regenerative Medicine and Biotherapy (IRMB) - Rapporteur
  • Vincent HUMBLOT - Directeur de recherche - Institut FEMTO-ST, Département Micro Nano Sciences & Systèmes - Rapporteur
  • Gaétan LAROCHE - Professeur - Université Laval - CoDirecteur de these
  • Corinne HOESLI - Associate Professor - McGill University - Examinateur
  • Frédéric HEIM - Professeur - Université de Haute Alsace - Examinateur

  Summary

  Engineering biomaterial surfaces functionalized with ligands derived from the extracellular matrix (ECM) is a multidisciplinary and expanding field essential for advances in regenerative medicine. Under physiological conditions, cells perceive multiple signals ranging from the nanometer to micrometer scale that guide their behavior and fate. The ECM, a dynamic and structured network, not only provides mechanical support but also delivers crucial biochemical and biophysical signals through its molecular architecture. Mesenchymal stem cells (MSCs), capable of self-renewal and multiple differentiation pathways, respond finely to their microenvironment. Their fate, particularly osteogenic differentiation, depends on factors such as substrate stiffness, nanotopography, and the presentation of bioactive motifs. Biomaterials can be optimized by incorporating adhesion peptides, growth factor mimetics, or ECM-inspired protein sequences. However, approaches limited to one or two peptides do not fully recapitulate the ECM's multifactorial complexity. Recent studies demonstrate that simultaneous presentation of multiple peptides induces synergistic effects on osteogenic marker expression. The success of these strategies relies on precise surface modification, with silanization playing a key role. Conventional methods such as solution immersion with commercial silanes suffer from variability that compromises reproducibility. To overcome these limitations, an innovative platform was developed on silicon wafers using spin coating combined with protected aminosilanes. This method produces robust and homogeneous self-assembled monolayers (SAMs) that allow controlled ligand presentation while demonstrating the critical influence of alkyl chain length and protective group chemistry on immobilization efficiency. Surface activation and silanization were characterized by contact angle measurements, PM-IRRAS, and XPS, establishing a reproducible protocol suitable for subsequent biological evaluations. On this basis, silicon surfaces were co-functionalized with up to three different peptides: RGD (integrin-mediated adhesion), P15 (collagen I mimetic), and BMP-2 (activation of osteogenic growth factor receptors). This novel combination recreates a complex and coherent biomimetic interface. Human MSCs cultured on these multifunctional substrates were analyzed by RT-qPCR and immunocytochemistry, revealing significant induction of early and late osteogenic markers, demonstrating accelerated and robust differentiation surpassing substrates functionalized with one or two peptides. In conclusion, this work represents a major advance in biomaterial surface engineering. It establishes that simultaneous co-functionalization with multiple peptides, based on the fine understanding of cell-ECM interactions, provides powerful and reproducible control over MSC osteogenic differentiation. This strategy paves the way for new biomimetic interfaces tailored for regenerative medicine applications.

• Formulation of an intermetallic matrix for the preparation of a composite with discontinuous reinforcement

  by Elena DAUFRESNE DE LA CHEVALERIE (Laboratoire des Composites ThermoStructuraux)

  The defense will take place at 10h00 - Amphithéâtre Laboratoire des Composites Thermostructuraux (LCTS) 3 Allée de la Boétie, 33600 Pessac

  in front of the jury composed of

  • Yann LEPETITCORPS - Professeur - Laboratoire des CompositesThermostructuraux, Université de Bordeaux - Directeur de these
  • Olivier DEZELLUS - Professeur - Laboratoire des Multimatériaux et Interfaces (LMI), Université Claude Bernard Lyon I - CoDirecteur de these
  • Gérard VIGNOLES - Professeur - Laboratoire des Composites Thermostructuraux, Université de Bordeaux - Examinateur
  • Aurélie JANKOWIAK - Ingénieure de recherche - ONERA - Examinateur
  • Jérôme CHEVALIER - Professeur - Laboratoire MatéIS, INSA Lyon - Rapporteur
  • Sylvain MARINEL - Professeur - Laboratoire CRISMAT, Université de Caen Normandie - Rapporteur

  Summary

  Ceramic Matrix Composites (CMC) are being investigated for the ‘ceramisation' of aircraft engines due to their ability to withstand higher temperatures than nickel-based superalloys currently in use. In particular, owing to their isotropic properties, short fibers reinforced CMC (SF-CMC) are considered for the manufacturing of small and complex-shaped parts at a lower cost. This thesis work is part of an industrial innovation approach with the main objective being the development of a new damage tolerant SF-CMC made up of an intermetallic matrix and Silicon carbide fibers Hi Nicalon type S as reinforcement. The Spark Plasma Sintering process was chosen to produce the composites, especially for its fast heating rates and its ability to improve the mechanical properties of the materials being investigated. Two matrixes were studied, both based on the MoSi2 compound which was chosen for its mechanical properties and resistance to oxidation at high temperatures and its thermochemical stability with SiC fibers. A first matrix (Mo,Ti)Si2-SiCp was synthesized by solid-state SPS from MoSi2, TiSi2 and SiC powders. The analysis of its microstructure and composition revealed the presence of secondary phases whose origin has been discussed. The particle size of the starting powder and the mixing process through powder grinding were found to significantly influence the complete conversion of the initial powders into the (Mo,Ti)Si2 solid solution. The poor mechanical properties of the latter are attributed to microcracking within the secondary phases and the inherently weak mechanical properties of the (Mo,Ti)Si2 phase. The addition of short SiC fibers at different volume ratio did not enable the composite to be damage tolerant, probably owing to the high thermal expansion of the matrix and its poor mechanical properties. Another matrix approach within the Mo-Si-Al-O-N chemical system was then examined. The selected formulation enabled a reduction in the matrix's thermal expansion while promoting a microstructural organization at the microscale. Its mechanical properties as well as those of the associated composite were significantly improved. Physicochemical analyses of the matrix highlighted the high sensitivity of its microstructure and composition to the temperature and pressure applied during the SPS process, owing to the presence of uncontrolled impurities. The use of coated short fibers with a double PyC-SiC layer resulted in a dense composite at high temperature while controlling the thickness of the mechanical interphase. However, the final developed SF-CMC is characterised by an early damage before its maximum stress and a brittle fracture, which raises questions about the value of adding short fibers. Optimisation strategies were explored and suggest that further improvements are possible in SiC fiber reinforced intermetallic matrix SF-CMC's mechanical properties. In conclusion, this work presents itself as a design guide for a new SF-CMC, from the matrix formulation to the characterisation of the final fiber reinforced material. It offers a methodological framework applicable to the innovation of a wide range of materials.

ED Sciences de la Vie et de la Santé

• Characterization of alterations in the splenic microenvironment of patients with myeloproliferative neoplasms

  by Bérénice DUGUE (Biologie des maladies cardiovasculaires)

  The defense will take place at 14h00 - Salle de Séminaire Inserm U1034 1 Av. Magellan, 33600 Pessac

  in front of the jury composed of

  • Chloé JAMES - Professeure des universités - praticienne hospitalière - Université de Bordeaux - Directeur de these
  • Lina BENAJIBA - Professeure des universités - praticienne hospitalière - Université Paris Cité - Rapporteur
  • David MICHONNEAU - Professeur des universités - praticien hospitalier - Université Paris Cité - Rapporteur
  • Diana PASSARO - Chargée de recherche - Université Paris Cité - Examinateur
  • Yacine BOULAFTALI - Chargé de recherche - Université Paris Cité - Examinateur
  • Damien LUQUE PAZ - Professeur des universités - praticien hospitalier - Université Angers - Examinateur

  Summary

  Myeloproliferative neoplasms (MPN) are chronic hematologic malignancies characterized by excessive blood cell production. Thrombosis represents the most common complication and the leading cause of mortality in these diseases. In MPN patients, thrombotic events may occur in unusual sites such as the splanchnic territory (which includes organs like the liver, the pancreas, and the spleen), an occurrence that is exceptionally rare in the general population. Interestingly, additional key aspects of MPN pathophysiology involve the spleen. This organ becomes a site of extramedullary hematopoiesis (EMH) developing in response to progressive bone marrow fibrosis and leading to splenomegaly. The development of a progressive fibrosis of variable intensity between individuals also characterizes spleens from MPN patients. Despite these major alterations, the splenic microenvironment has been comparatively understudied compared to the bone marrow niche. We hypothesize that its remodeling in patients with myeloproliferative neoplasms (MPN) contributes to the increased risk of splanchnic thrombosis by promoting activation of splenic endothelial cells. Furthermore, since endothelial cells can acquire a pro-fibrotic phenotype in various pathological contexts, we also hypothesize that they contribute, in collaboration with other surrounding cell types of the microenvironment, to the development of splenic fibrosis. To investigate these mechanisms, we performed spatial transcriptomics on 12 human spleen FFPE samples, including 9 from MPN patients and 3 from healthy controls. We also analyzed 8 of these samples using snRNAseq. By integrating both datasets, we generated the first comprehensive spatial transcriptomic atlas of the healthy and MPN human spleen at single-cell resolution. At the transcriptomic level, we show that disorganization of the splenic architecture (including red pulp hyperplasia, EMH, and reduced white pulp) is accompanied by extensive transcriptomic remodeling of the splenic microenvironment. In particular, splenic endothelial cells emerge as key players, adopting a pro-adhesive and pro-thrombotic phenotype that may locally contribute to the increased thrombotic risk of patients. We also highlight an important reprogramming of fibroblastic reticular cells. In collaboration with Prof. Schneider's group, we demonstrated that these alterations are faithfully recapitulated in two murine MPN models representing moderate (ThPO overexpression) and severe (MPL_W515L overexpression) myelofibrosis, studied by scRNAseq and multiplex CODEX imaging. At early disease stages, these cells appear to exert a supportive function for EMH, before shifting at later stages towards a pro-inflammatory and pro-fibrotic phenotype. Importantly, our results also reveal an activation of the complement system, primarily mediated by splenic reticular cells and macrophages. Finally, we show that in vivo complement inhibition in the non-hematopoietic compartment of MPN mice reduces both splenomegaly and bone marrow fibrosis, while normalizing hematologic parameters. This demonstrates that the complement axis represents a global therapeutic target. Together, our findings underscore the underappreciated contribution of the splenic microenvironment to both thrombotic and fibrotic complications in MPN patients.

ED Sciences Physiques et de l'Ingénieur

• Influence of human–machine cooperation mode on occurrent representation in a dynamic situation and cognitive load: the case of swarm management by fighter pilots

  by Benjamin COULOMB (Laboratoire de l'Intégration du Matériau au Système)

  The defense will take place at 10h00 - O108 école nationale supérieure de cognitique 109 Av. Roul, 33400 Talence,

  in front of the jury composed of

  • Jean-Marc ANDRE - Professeur des universités - Université de Bordeaux - Directeur de these
  • Françoise DARSES - Directrice de recherche - Institut de recherche biomédical des armées - CoDirecteur de these
  • Philippe RAUFFET - Professeur des universités - Université de Bretagne Sud - Examinateur
  • Julien CEGARRA - Professeur des universités - Institut national universitaire Champollion - Examinateur
  • Marie-Pierre PACAUX-LEMOINE - Directeur de recherche - Université Polytechnique Hauts-de-France - Rapporteur
  • Jordan NAVARRO - Professeur des universités - université lumière Lyon 2 - Rapporteur

  Summary

  This doctoral research investigates the cognitive constraints faced by fighter pilots engaged in the supervision of an automated weapon system (the swarm) operating in dynamic, complex, and hostile environments. The ultimate aim is to provide anthropocentric design orientations for the swarm, not only like a tool, but as an artificial agent capable of active cooperation with pilots. The activity under study concerns the post-release phase, i.e., the period between the weapon launch and the confirmation of target destruction. To accomplish their mission during this phase, pilots must simultaneously supervise the weapon system and essential survival processes (e.g., flight path, fuel management, radar). The introduction of the swarm, enabled by artificial intelligence technologies, profoundly reshapes this activity. With its high level of automation, the swarm shifts from the status of tool to that of teammate, thereby requiring cooperative interaction with the pilot. It thus becomes crucial to ensure an interaction that both preserves the quality of the occurrent representation and maintains an appropriate level of cognitive load. To account for operator–system interactions, this research is grounded in the theoretical framework of human–machine cooperation. This field highlights the importance of an appropriate function allocation between human and machine to optimise collective performance. Within this framework, the thesis relies on four cooperation modes to define alternative function allocations between the operator and the swarm: (1) Perceptive; (2) Mutual control; (3) Function delegation; (4) Automated. These modes correspond to increasing levels of automation of the swarm. The research was structured into two empirical studies. The first study aimed to identify a support task suitable for the subsequent evaluation of the cooperation modes. For this purpose, evaluation criteria were developed, encompassing both cognitive and operational dimensions of activity. Twenty-one operational fighter pilots were recruited to assess eight swarm management tasks. The results led to the identification of four priority tasks, among which “trajectory modification” was retained as the support task. The second study examined the impact of cooperation modes on cognitive load and the precision of the occurrent representation. Twenty-eight Rafale and Mirage 2000D pilots participated. The findings revealed that cognitive load decreases in the more automated modes (#3 and #4). However, only in mode #3 (“function delegation”) did the precision of the occurrent representation significantly decline. This effect can be explained by the marked reduction in diagnostic activities, evidenced through self-confrontation interviews conducted after the experimental sessions. Analysis of these interviews further revealed two distinct cognitive strategies as a function of automation level: the manual modes (#1 and #2) induced an “anticipatory” strategy, whereas the automated modes (#3 and #4) fostered a “reactive” strategy. In conclusion, the thesis contributes to (1) advancing understanding of future swarm management activities; (2) quantifying the effects of different function allocations within a weapon system on cognitive load and occurrent representation; and (3) providing insights into the cognitive strategies adopted by pilots in the management of dynamic situations. These findings also yield valuable design implications for the development of future cooperative weapon systems.

• Light field microscopy with enhanced spatial bandwidth product for the observation of organoids

  by Aymerick BAZIN (Laboratoire Photonique, Numérique & Nanosciences)

  The defense will take place at 14h00 - Amphitheatre André Ducasse Institut Optique d'Aquitaine 1 Rue François Mitterrand, 33400 Talence

  in front of the jury composed of

  • Amaury BADON - Chargé de recherche - Université de Bordeaux - Directeur de these
  • Corinne LORENZO - Ingénieur de recherche - Institut des Technologies Avancées en sciences du Vivant - Rapporteur
  • Loic LE GOFF - Directeur de recherche - Institut Fresnel - Rapporteur
  • Vincent STUDER - Directeur de recherche - Institut Interdisciplinaire de Neurosciences (IINS) de Bordeaux - Examinateur
  • Antoine AUBRET - Chargé de recherche - Laboratoire Ondes et Matière d'Aquitaine - Examinateur
  • Giovanni CAPPELLO - Directeur de recherche - Laboratoire Interdisciplinaire de Physique, Université Grenoble Alpes - Examinateur

  Summary

  Volumetric imaging is essential for the non-invasive observation of biological samples or industrial components. Such methods must possess optical sectioning capabilities, collecting light only from a specific point or plane while rejecting signals from other depths. Volume acquisition is typically achieved by axially translating the sample. However, this approach can be problematic for biological systems immersed in aqueous media, which cannot remain perfectly still during sample displacement. This is particularly true for organoids, which are multicellular aggregates that mimic certain organ functions in vitro. In this context, light field microscopy (LFM) emerges as a promising alternative, as it enables instantaneous 3D imaging through the insertion of a microlens array in the image plane of a microscope. This array captures angular information at each point of the sample; processing this 4D image then allows for the extraction of volumetric data. However, LFM faces a major limitation: the trade-off between spatial and angular resolution, due to the limited number of pixels on the sensors. The objective of this thesis is to overcome these limitations by developing an improved version of LFM, while increasing the spatial bandwidth product. This goal will be pursued through three specific objectives: (1) increase the amount of information captured in a single acquisition by replacing sample displacement with a remote scanning system, thereby expanding the field of view and improving spatial resolution; (2) develop a compact version of the system compatible with integration into an incubator; and (3) enable long-term imaging of organoids (over 15 days).

ED Sciences et environnements

• Endemic fluorosis in the Roman Era within the campanian volcanic Arc (Cumae and Pompeii): a textbook case in historical ecotoxicology

  by Eliza ORELLANA GONZÁLEZ (De la Préhistoire à l'Actuel : Culture, Environnement, Anthropologie)

  The defense will take place at 14h00 - Amphithéâtre GABA Amphithéâtre GABA (bâtiment B5, campus de Talence) 33400 Talence

  in front of the jury composed of

  • Dominique CASTEX - Directrice de recherche - CNRS - Université de Bordeaux - Directeur de these
  • Caroline POLET - Professeur chargé de cours - Institut royal des Sciences naturelles de Belgique - Rapporteur
  • Nick SCHIAVON - Directeur de recherche - University of Évora - Rapporteur
  • Sacha KACKI - Chargé de recherche - CNRS - Université de Bordeaux - Examinateur
  • Nicolas LAUBRY - Maître de conférences - Université Paris-Est Créteil - Examinateur
  • Rémy CHAPOULIE - Professeur des universités - Université Bordeaux Montaigne - CoDirecteur de these

  Summary

  This research explores the health impact of chronic fluoride exposure in Antiquity, through the study of the populations of Cumae and Pompeii, located within the Campanian volcanic arc. While fluoride is essential for bone and dental mineralization, excessive or chronic intake can cause skeletal fluorosis, a condition characterized by hyperostosis, osteosclerosis, and sometimes soft-tissue ossifications. In volcanic areas, where fluoride is naturally present in soils and waters, this condition can become endemic. The funerary contexts examined (2nd c. BCE – 1st c. CE) consist of secondary cremation burials, a dominant practice in Republican and Imperial Rome. These remains, altered by fire, pose significant diagnostic challenges: absent or highly fragmented teeth, fragmented bones limiting the observation of macroscopic lesions, and chemical transformations of bone. Despite these constraints, preliminary anthropological observations at Cumae revealed unusually high bone mass suggestive of a metabolic disorder, whereas such a signal was not observed at Pompeii. The research employed an interdisciplinary approach combining paleopathology, archaeometry, and funerary archaeology. A protocol for macroscopic observations was established across twelve anatomical zones to record lesions consistent with fluorosis. In parallel, Laser-Induced Breakdown Spectroscopy (LIBS) was applied to detect fluoride in burned bones, an innovative method validated through comparison with reference analyses (PIXE, PIGE). This combination demonstrated the feasibility of fluoride detection in such materials and led to the proposal of an indicative threshold (0.095 a.u. LIBS) for suspecting fluorosis in comparable archaeological contexts. The results reveal a clear difference between the two sites: at Cumae, a high frequency of bone lesions and elevated fluoride levels suggest significant chronic exposure, linked to the proximity of the Phlegraean Fields caldera and a probably subterranean water supply. At Pompeii, contamination appears more diffuse, associated with distinct water systems (aqueduct, cisterns, Somma-Vesuvius aquifers). A diachronic analysis, covering the period represented by the studied corpus, also highlights variations in exposure intensity over time. Beyond paleopathological diagnosis, this study contributes to a broader reflection within the emerging field of historical ecotoxicology, which examines the delayed and cumulative effects of environmental pollutants on past societies. By integrating bioarchaeological, archaeometric, and historical data, it sheds new light on health risk factors among ancient populations and invites us to view volcanic environments not only as resources but also as potentially toxic settings. This work thus helps to lay the methodological and theoretical foundations of historical ecotoxicology, attentive to processes of chronic exposure, bodily incorporation, and physiological adaptation in ancient populations.