I am an Experimental Physicist and Machine Learning practitioner with more than 20 years of experience in quantitative scientific research. My work has ranged from designing new experimental approaches to investigate complex physical systems to developing and systematically evaluating modern Machine Learning methods in physics, medical imaging, computational biology and real-world Data Science.
A common thread throughout this work is the formulation of difficult quantitative problems in ways that allow competing mechanisms, models or explanations to be tested systematically. In recent years, I have increasingly applied this scientific methodology using modern Machine Learning and Deep Learning methods.
Selected Scientific Research and Machine Learning · Machine Learning Projects · Experimental & Scientific Leadership · Research Profiles
Proposed and led a new experimental approach using polarization observables to investigate possible in-medium modifications of baryon resonances in heavy nuclei at the University of Mainz. The central idea was to use beam-helicity asymmetries as a probe potentially less affected by final-state interactions than conventional unpolarized observables, allowing a cleaner investigation of changes in the underlying reaction dynamics.
Main contributions:
Outcome: First measurements of beam-helicity asymmetries for π⁰η photoproduction on carbon, aluminum and lead, demonstrating the feasibility of this approach and opening a route for further studies of in-medium effects. First-author publications in Physics Letters B and Physical Review C.
Links:
Developed an event-by-event Machine Learning method for identifying time-correlated physical events in tagged-photon experiments as an alternative to conventional statistical background subtraction. The approach combines GEANT4-simulated signal events with experimentally measured background and was validated against established physics-analysis methods.
Main contributions:
Outcome: The Machine Learning approach reproduced the results of the conventional physics analysis while enabling event-by-event identification of correlated events. Published in the Journal of Instrumentation (JINST).
Link:
Developed and carried out an alternative experimental approach to constraining the proton scalar polarizabilities, fundamental quantities describing the response of the nucleon to electromagnetic fields, in research at George Washington University and the University of Mainz. The project used the beam asymmetry Σ₃ in Compton scattering as a complementary observable to conventional unpolarized cross-section measurements.
Main contributions:
Outcome: First measurement of the beam asymmetry Σ₃ in proton Compton scattering below pion-production threshold, establishing an alternative experimental route for constraining the proton scalar polarizabilities.
Link:
Investigation of the excitation spectrum and internal dynamics of the nucleon through double-meson photoproduction with the CBELSA/TAPS experiment at ELSA, University of Bonn. The work used polarization observables and multidimensional analysis of complex three-body final states to disentangle contributions from overlapping baryon resonances and competing sequential decay mechanisms.
Main contributions:
Outcome: The measurements provided new constraints on the nucleon excitation spectrum, contributed to the identification of previously unobserved resonance decay modes, and provided new information on the internal structure of excited baryon states. One resulting collaborative publication (Eur. Phys. J. A 51, 95 (2015)) was selected by the Particle Data Group (PDG) as the primary reference for properties of multiple baryon resonances.
Links:
Development of a solution for multi-class chest X-ray abnormality detection, achieving 3rd place on the private leaderboard among 200+ participants.
The project systematically investigates detector complementarity across architectures (RT-DETR, DINO), optimization strategies (ADAM, MuSGD), and cross-validation folds. The central finding is that selective evidence fusion, using independent detectors as supporting evidence rather than fusing all predictions, consistently outperforms conventional ensembling approaches.
Main topics:
Outcome: 3rd place (private leaderboard). Comprehensive technical writeup published.
Links:
Development and systematic evaluation of Deep Learning methods for automated classification of dopamine transporter (DaT) SPECT scans into normal and pathological cases using a multicenter dataset collected across ten French hospitals. The project investigates custom 3D convolutional neural networks and pretrained medical vision foundation models, including MedSigLIP, using zero-shot inference, frozen feature extraction, parameter-efficient adaptation and supervised fine-tuning. The work focuses on controlled comparison of modelling approaches, transfer learning, augmentation strategies, generalization across heterogeneous imaging data, failure analysis and reproducible PyTorch-based experimentation.
The challenge is organized by the French Society of Nuclear Medicine (SFMN) in partnership with the Health Data Hub and GaelO as part of the France 2030 Health Data Challenges initiative. Final results and technical analysis will be published after completion of the project.
Competition: https://www.drivendata.org/competitions/311/dat-parkinsons-challenge/
Development of transformer-based methods for automatic protein function annotation using ProtT5-XL and ESM2 foundation models, including end-to-end fine-tuning and systematic benchmarking against fixed-embedding baselines.
A key methodological contribution is an information accretion-aware ensemble aggregation strategy, which uses different aggregation schemes for ontology regions depending on label specificity, addressing the challenge of rare-label signal preservation in hierarchical multi-label prediction.
Main topics:
Outcome: Detailed methodological report published. Work in progress toward a scientific publication.
Links:
Development and integration of active time-projection chamber systems for precision proton-radius and polarizability measurements at the University of Mainz.
Main contributions:
Scientific referee for the majority of publications of the A2 Collaboration at MAMI (2018–2023). Responsibilities included critical review of experimental analyses, assessment of systematic uncertainties and error propagation, reproducibility and consistency checks, and scientific quality assurance prior to publication.