ASNR / Mines Paris PSL
Oct 01, 2026 (First Call)
Context
On nuclear sites, controlling atmospheric contamination is critical to protect workers from radiation.
Continuous Air Monitors (CAMs) are deployed to continuously collect airborne aerosols on a filter,
assess the accumulated radioactivity in real-time, and trigger alerts if contamination is detected [13].
However, in decommissioning sites, these instruments may exhibit unexpected behavior due to specific
atmospheric conditions, such as dust from concrete scraping or steel cutting [7]. The deterministic
algorithm currently implemented in these instruments, which decides whether to trigger an alarm
based on the recorded spectrum, is not always suitable for such non-standard atmospheric conditions
defined by the International Electrotechnical Commission (IEC 60761 series).
To improve the performance of current instruments, the French Authority for Nuclear Safety and
Radiation Protection (ASNR) in partnership with Mines Paris-PSL has recently developed a novel
Deep Learning-based algorithm [10, 11]. This algorithm is effective in dusty conditions and achieves
the same level of accuracy as the current algorithm in normal conditions. With the emergence of
AI systems, a new research field focuses on analyzing confidence in their predictions to ensure an
operational use of AI in industrial applications. This confidence is particularly important when the
algorithm supports expertise and decision-making, as in nuclear safety and radiological protection.
The concept of confidence has already been formalized in an industrial context by the European Union
Aviation Safety Agency [6] and more recently by ASNR, EDF R&D, Framatome, and CEA for nuclear
activities [5], based on five guiding principles: consistency, exactness, robustness, intelligibility, and
transparency.
Position Overview
This postdoctoral position, proposed by ASNR and MINES Paris, is included in the EXODD project in
partnership with EDF R&D with the objective to apply these guiding principles to the Deep Learning
algorithmforspectrumanalysis. Thegoalistoaddressopenresearchquestionsrelatedtoexactnessand
robustness to enable its operational implementation in measurement instruments and its application
to real spectra. The candidate is invited to present his or her findings at international conferences.
The work will focus on the following three development axes:
Axis 1: Improving Model Performance with High-Fidelity Data
This axis focuses on the principle of exactness, ensuring acceptable precision in AI system predictions.
So far, the spectrum analysis algorithm has been trained, validated, and tested on a synthetic database.
To guarantee acceptable model performance in real-world situations, ASNR is currently developing a
high-fidelity simulation code that more precisely describes aerosol deposition mechanisms, albeit at a
highcomputationalcost[1]. Theobjectiveofthisaxisistostudyhowhigh-fidelitydatacanimprovethe
algorithm’s real-world performance by combining synthetic and high-fidelity data via transfer learning
[9, 3] and/or fine-tuning, while using active learning [12] and refinement criteria [4] to strategically
enrich the limited dataset.
Axis 2: Operational Exploitation of the Model
This axis focuses on the principle of robustness, evaluating the AI system’s performance in the presence
of variability. To address this, the work will focus on defining the validity domain of the model, and/or
developing a tool to detect whether a new real spectrum significantly differs from the spectra in the
training database (in the sense of distance in the latent space). Additionally, building on the work
in [11] on uncertainty quantification, conformal prediction will be used to construct OOD detection
scores, which is an active research area in the AI community.
Axis 3: Operational Implementation in Measurement Instruments
This axis exploits the contributions of the two previous axes for real-world deployment. The designed
modelwillbeembeddedonmicrocontrollers, withoptimizationstoreducememoryfootprintandenergy
consumption. Tests will be conducted on ICARE [2] to validate the system’s behavior under realistic
conditions [8]. Finally, an automatic anomaly diagnosis system will be integrated to detect hardware
failures or sensor drifts.
Requirements
• Ph.D. in Statistics, Machine Learning, Optimization or Signal/Image Processing.
• Strong background in Machine/Deep Learning and probabilistic modeling.
• Experience with Python and data analysis tools.
• Knowledge of nuclear physics or radiological protection is a plus.
• Knowlegde of uncertainty quantification is a plus.
• Fluency in English is required; proficiency in French is a plus.
Duration and Location
• Duration: 18 months.
• Location: ASNR, Saclay, France.
• Deadline for applications: (31 December, 2026) (Open until filled)
• Starting day: The start date can be adjusted to accommodate the candidate.
Application
Interested candidates should submit:
• A detailed CV.
• Contact information for at least two references.
• List of publications.
Contact:
For further information, please contact:
• Santiago VELASCO-FORERO santiago.velasco@mines-paristech.fr
• Jean BACCOU jean.baccou@asnr.fr
• Grégoire DOUGNIAUX gregoire.dougniaux@asnr.fr
References
[1] S. Agostinelli, Geant4—a simulation toolkit. Nuclear Instruments and Methods in Physics
Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment,506(3):250–
303, 7 2003.
[2] M. Ammerich. Réalisation d’une installation d’étalonnage de moniteurs de contamination at-
mosphérique à l’aide d’aérosols radioactifs calibrés (ICARE). Technical report, Conservatoire
National des Arts et Métiers, Fontenay-aux-Roses, 1989.
[3] J. J. Bird, D. R. Faria, A. Ekárt, and P. P. S. Ayrosa. From simulation to reality: CNN trans-
fer learning for scene classification. In 2020 IEEE 10th International Conference on Intelligent
Systems, pages 619–625, 2020.
[4] W. Cai, M. Zhang, and Y. Zhang. Batch mode active learning for regression with expected model
change. IEEE Transactions on Neural Networks and Learning Systems, 28(7):1668–1681, 2017.
[5] E. Chojnacki and S. Bannouf. Présentation IA de confiance. GT IA COFREND, www.cofrend.
com, 2025.
[6] EASA. Concept Paper: Guidance for Level 1 & 2 ML applications. Technical report, European
Union Aviation Safety Agency, 2024.
[7] G. Hoarau. Etude de la limite de détection et des fausses alarmes émises par les moniteurs de
mesure de la contamination radioactive atmosphérique dans les chantiers de démantèlement. PhD
thesis, Université Paris-Saclay, 2020.
[8] C. Monsanglant-Louvet, N. Liatimi, and F. Gensdarmes. Performance Assessment on Contin-
uous Air Monitors Under Real Operating Conditions. IEEE Transactions on Nuclear Science,
59(4):1414–1420, 8 2012.
[9] S. J. Pan and Q. Yang. A survey on transfer learning. IEEE Transactions on Knowledge and
Data Engineering, 22(10):1345–1359, 2010.
[10] A. Roblin, J. Baccou, G. Dougniaux, and S. Velasco-Forero. Deep learning approach for airborne
alpha radioactivity monitoring in atypical atmospheric conditions. Journal of Aerosol Science,
187:106573, 2025.
[11] A. Roblin, J. Baccou, G. Dougniaux, and S. Velasco-Forero. Uncertainty quantification in ra-
dioactivity level estimation via heteroscedastic count regression. In 2026 IEEE World Congress
on Computational Intelligence (WCCI 2026), IEEE Computational Intelligence Society (CIS),
2026 (Oral Presentation).
[12] B. Settles. Active learning literature survey. University of Wisconsin-Madison Department of
Computer Sciences, 2009.
[13] L. Zhengyong and J. J. Whicker. Considerations for data processing by continuous air monitors
based on accumulation sampling techniques. Health Physics, 94(2), 2008.
