Optimization of a metrological approach of fast spectroscopic identification of g-emitting radionuclides at low statistics
Keywords: automatic identification, spectral unmixing, machine learning, nuclear instrumentation, spectroscopic measurements
Context: Fast and robust spectroscopic identification of g-emitting radionuclides from portal radiation monitors (e.g. located at borders, airports, harbors, etc) is key to prevent from illegal nuclear material trafficking. In this context, the development of reliable automatic identification systems is mandatory to minimize the number of expert interventions in decision-making due to the continuous flow of persons, vehicles, luggage, etc. In the framework of the French ANR project NANTISTA, the LNE-LNHB (Laboratoire national Henri Becquerel) located at CEA/Saclay (France) has developed an automatic identification algorithm based on spectral unmixing for fast anomaly detection of g-emitting radionuclides in natural background radiation. This identification code was first designed to be implemented in portal radiation monitors equipped with scintillation detectors (plastic, NaI(Tl)). The goal was also to meet a strong need for radionuclide identification systems in the field of environmental measurements. In particular, nuclear instrumentations used for radiological or nuclear accident by IRSN (Institut de radioprotection et de sûreté nucléaire), which is in charge of nuclear safety in France (www.irsn.fr/EN/).
The objective of the post-doctoral position is to optimize the spectral unmixing and decision-making algorithms in order to achieve low false alarm rates at low statistics in accordance with current standards and guidances (ISO 22188:2004, IAEA NSS1, 2010, ANSI N42.38-2006). For that purpose, recent advances in machine learning will be investigated in collaboration with CEA/LCS (Laboratoire de cosmologie et statistique) located at CEA/Saclay. In particular, these investigations will deal with the improvements of: i) decision making at low statistics; ii) the robustness of the spectral unmixing regarding overfitting in the case of mixtures of several g-emitting radionuclides having overlapping spectral signatures. A potential application of the improved identification algorithm can be its implementation on a portable instrumentation to be applied by non-expert staff in the case of in-situ environmental analysis following a radiological or nuclear accident. This work will be carried out in collaboration with the Laboratoire de mesures nucléaires at IRSN.
The candidate must have a PhD in machine learning and/or signal processing. Experience in probabilistic programming and anomaly detection is also welcome. Additionally, excellent programming skills are required (Python, C++),
The offer is a full-time position for initially 12 months at CEA/LNHB (www.lnhb.fr/en/). The LNE-LNHB is in charge of French references in the field of metrology of ionizing radiation. Regarding investigations related to machine learning, the post-doctoral work will be conducted in collaboration with CEA/LCS (www.cosmostat.org).
Interested candidates should send a CV, containing a list of publications and a summary of past research.
Contact: Christophe BOBIN (CEA/LNHB),email@example.com
(c) GdR 720 ISIS - CNRS - 2011-2018.