Click here to

Session: Multi-Disciplinary General ePoster Viewing [Return to Session]

RAIdiation Protection Co-Pilot:A Machine Learning Aided Radiation Scenario Prediction Tool

E Derugin1*, K Kroeninger1, O Nackenhorst1, F Mentzel1, J Walbersloh2, J Weingarten1, (1) Department of Physics, TU Dortmund University, Dortmund, NW, DE, (2) Materialpruefungsamt Nordrhein-Westfalen, Dortmund, NW, DE

Presentations

PO-GePV-M-36 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: The Department of Physics at the TU-Dortmund University, in cooperation with the Materialprüfungsamt Nordrhein-Westfalen (MPA NRW) is developing multivariate glow curve analysis techniques based on the new thermoluminescence dosimeter system TL-DOS using convolutional neural networks (CNN), which makes it possible to retrospectively obtain additional knowledge about the irradiation beyond the irradiation dose estimation. Information about the day of irradiation or the number of irradiation fractions can be used to trace the circumstances of exposure and thus improve the existing radiation protection concept. First proof-of-concept studies of the prediction of the irradiation date were successfully performed on data sets, which only consider the variation of the irradiation date for the same irradiation dose and are therefore limited in their transferability to real-world application. [F. Mentzel et. al. 2019 J. Rad. Meas. 125(10), F. Mentzel et al 2021 J. Radiol. Prot. 41 S506]

Methods: Several thousands of glow curves were measured to create a data set that is comparable to a real-world application with different doses and storage conditions. Furthermore, the detectors are preheated before the measurement as in routine dosimetry of the MPA NRW. We train a CNN with this data set and compare the prediction accuracy with the results of the proof-of-concept studies. The advantage of using CNN is that the required glow curves hardly need to be preprocessed, thus simplifying the information extraction process.

Results: We achieve a prediction accuracy of up to 4 days for the estimation of the irradiation day for single-dose irradiation of 10mGy within a monitoring interval.

Conclusion: To integrate neural network predictions in a routine personal dosimetry setting, we present a prototype ML-based prediction tool that provides information about the irradiation day as well as the irradiation dose and the interpretability of the prediction.

Funding Support, Disclosures, and Conflict of Interest: This research was supported by the Deutsche Forschungsgemeinschaft (DFG), Project No. KR 4060/10-1.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Share: