Thèse Potentiels d'Apprentissage Automatique pour la Réactivité Radicalaire aux Surfaces de Nanoparticules H/F - Doctorat.Gouv.Fr

Établissement : Université Paris-Saclay GS Chimie École doctorale : Sciences Chimiques : Molécules, Matériaux, Instrumentation et Biosystèmes Laboratoire de recherche : Institut de Chimie Physique Direction de la thèse : Van Oanh NGUYEN THI ORCID 0000000170596696 Début de la thèse : 2026-09-01 Date limite de candidature : 2026-06-30T23:59:59 La radiothérapie repose sur la génération de radicaux hydroxyle (HO-) par les rayonnements ionisants pour endommager l'ADN des cellules tumorales. Certaines nanoparticules, notamment les nanoparticules d'or (AuNPs) et les nanostructures carbonées, agissent comme des amplificateurs locaux de radiation en augmentant fortement la production de radicaux à leur surface. Cependant, les mécanismes moléculaires à l'origine de ce phénomène restent mal compris. Ce projet de thèse, s'inscrivant dans le projet RAINHET (Réactivité Radicalaire aux Interfaces de Nanostructures Hybrides : Synergie Expérience-Théorie), vise à combler cette lacune par une approche de modélisation computationnelle avancée. Le/la candidat(e) développera des potentiels interatomiques par apprentissage automatique (MLIPs), entraînés sur des données de théorie de la fonctionnelle de la densité (DFT), pour simuler avec une précision quasi-quantique la structure et la réactivité radicalaire des surfaces de nanoparticules en milieu aqueux. Des simulations de dynamique moléculaire à grande échelle seront réalisées sur des modèles réalistes de nanoparticules afin d'étudier l'organisation de l'eau et la formation de HO-. Des spectres SFG (génération de fréquence somme) simulés seront calculés à partir des trajectoires pour permettre une comparaison directe avec les données expérimentales. Ce travail est effectué à l'Institut de Chimie Physique et s'inscrit dans le cadre d'un programme de thèse en cotutelle entre l'Université Paris-Saclay (France) et l'Université des Sciences de l'Université Nationale du Vietnam à Hanoï (VNU-HUS). Radiotherapy is one of the most widely used treatments for cancer. It works by exposing tumors to ionizing radiation, which breaks apart water molecules and generates highly reactive chemical species, in particular, hydroxyl radicals (HO-). These radicals damage the DNA of cancer cells and trigger cell death. A major challenge in radiotherapy is limiting collateral damage to healthy tissues surrounding the tumor. To address this, researchers have explored using nanoparticles (NPs) as local radiation enhancers: when irradiated, certain nanoparticles dramatically amplify the local production of HO- radicals, concentrating the therapeutic effect precisely where needed. Two types of nanoparticles are of particular interest in this project: (i) Gold nanoparticles (AuNPs), which are known to massively amplify HO- production under irradiation. (2) carbon-based nanoparticles that show similar radical-enhancing behavior and offer rich surface chemistry.

Despite experimental evidence of enhanced radical production, the molecular mechanisms behind this phenomenon remain poorly understood. Key open questions include: (i) How exactly does the nanoparticle surface interact with surrounding water to generate HO- radicals? (ii) How does functionalizing the nanoparticle surface influence radical production? (iii) Can we design hybrid nanostructures that combine the advantages of both materials? Answering these questions requires a tight coupling between state-of-the-art experiments and advanced computational simulations, which is precisely the ambition of the RAINHET project.
RAINHET (Radical Reactivity at Hybrid Nanostructure Interfaces: Experiment-Theory Synergy) brings together three research groups at the Institut de Chimie Physique (ICP, Paris-Saclay) with complementary expertise in nanoparticle synthesis, Sum-Frequency Generation (SFG) nonlinear optical spectroscopy, and computational chemistry. The computational modeling aspect will consist of developing machine learning interatomic potentials (MLIPs) to simulate the structure and reactivity of nanoparticle surfaces at the atomic scale, and use these to predict and interpret experimental SFG spectra.

Simulating nanoparticles at atomic resolution is computationally very demanding. Traditional quantum chemistry methods are too slow for particles of realistic size (3-10 nm in diameter, containing thousands of atoms). Machine learning interatomic potentials (MLIPs) offer a compelling solution: they are trained on quantum chemistry data and can achieve near-quantum accuracy at a fraction of the computational cost. This makes it possible to run long molecular dynamics simulations on realistic nanoparticle models, and to compute properties such as vibrational spectra directly comparable to SFG experiments.

Scientific objectives
- Obj. 1 Build training datasets from quantum chemistry (DFT) calculations on model nanoparticle surfaces, including water molecules and radical species.
- Obj. 2 Train machine learning potentials (deep neural network potentials, message-passing networks, or transformer-based models) on these datasets.
- Obj. 3 Use active learning strategies to efficiently expand the training set and improve model transferability, especially for reactive configurations.
- Obj. 4 Run large-scale molecular dynamics simulations to study water organization and radical formation at the surface of AuNPs, carbon-based NPs, and hybrid nanoparticles.
- Obj. 5 Compute simulated SFG spectra from the trajectories (via dipole moments and polarizability tensors) to compare directly with experiments.

We are looking for motivated candidates with a Master's degree (or equivalent) in one of the following fields:
- Theoretical or Computational Physics
- Computational Materials Science or Physical Chemistry

• Chimie