Machine learning and multimethod-NDE for estimating neutron-induced embrittlement

Jari Rinta-aho, Sonja Grönroos

Research output: Chapter in Book/Report/Conference proceedingConference abstract in proceedingsScientific

40 Downloads (Pure)


In this study, 157 irradiated and non-irradiated Charpy specimens
[1] manufactured from six different steel alloys used in the reactor
pressure vessels (18MND5-W, 22NiMoCr37, A508-B, 15Kh2NMFA,
HSST03 and A508-Cl2) were measured. The measurements included
determining several non-destructively measurable electric, magnetic
and elastic parameters. The applied non-destructive methods were
Direct Current-Reversal Potential Drop (resistivity) [2], 3MA (eddy
current impedance loop shape) [3], TEP (Seebeck Coefficient) [4], MIRBE (Barkhausen noise) [5], MAT (magnetic hysteresis loop shape) [5] and sound velocity. After the non-destructive measurements, the ductile-brittle transition temperature (DBTT) was determined destructively using the ISO-standard method [1]. Several different regression algorithms, including neural network regression and support vector regression, were applied to the data. The algorithms were implemented with TensorFlow and scikit-learn using Python 3.7. With these algorithms, it was possible to estimate the DBTT with the mean absolute error smaller than 20 °C. Based on the results, the method can be seen as a potential candidate for estimating neutron-induced embrittlement non-destructively.
Original languageEnglish
Title of host publicationEuropean NDE Symposium for NPP
Subtitle of host publicationEvent booklet
PublisherNOMAD project
Number of pages1
Publication statusPublished - 5 May 2021
MoE publication typeNot Eligible
EventEuropean Symposium on NDE for NPP - Franhofer Institute, Saarbruchen, Germany
Duration: 4 May 20215 May 2021


ConferenceEuropean Symposium on NDE for NPP


Dive into the research topics of 'Machine learning and multimethod-NDE for estimating neutron-induced embrittlement'. Together they form a unique fingerprint.

Cite this