E L E C T R O N I C S T R U C T U R E , M A G N E T I S M A N D D Y N A M I C S

S C I E N T I F I C H I G H L I G H T S

1 2 2 H I G H L I G H T S 2 0 2 1 I

Neural network analysis of EXAFS spectra to identify technetium chukanovite species

Machine learning was applied to analyse EXAFS spectra of chukanovite (Fe2(OH)2CO3), identified as an efficient reductant and immobilisation agent for nuclear fission product technetium (Tc). Information on spectral endmembers and their dependence on geochemical parameters is needed to develop sound thermodynamic models to predict the safety of nuclear waste repositories.

The fission product 99Tc plays a crucial role in release scenarios for safety assessments of repositories for high- level radioactive waste. Tc can occur in oxidation states ranging from -I to +VII [1]. Its environmental mobility is mainly determined by redox conditions. Under oxic conditions, the very mobile Tc(VII)O4

− is predominant. Under anoxic conditions, however, interaction with Fe(II)- bearing solids commonly present as steel corrosion products but also as natural minerals, leads to formation of Tc(IV), which is efficiently immobilised by sorption to and/ or incorporation by primary and secondary mineral phases [2,3].

Chukanovite (Fe2(OH)2CO3), a relevant corrosion product of carbon steel under nuclear waste repository conditions, has rarely been investigated because of the difficulty in conserving its pristine Fe(II) state in the lab. By using

an elaborate system of oxygen-free alpha-gloveboxes, sample shipment under LN2 and EXAFS measurements in a He cryostat at beamline BM20, it was possible to obtain EXAFS spectra of 37 samples representative of the anoxic conditions prevailing in deep geological formations. Using Tc K-edge EXAFS spectroscopy, two series of samples were investigated, Tc chukanovite sorption samples and coprecipitates, prepared under varying geochemical conditions. The retention potential of chukanovite towards Tc(VII) was found to be high in the pH range 7.8 to 12.6, evidenced by high solid-water distribution coefficients, log Rd ~ 6.

From these 37 EXAFS spectra, spectral endmembers and their dependence on geochemical parameters were derived by self-organising (Kohonen) mapping (SOM), a biologically inspired neural network-based machine learning process [4]. SOM is used to explore the hidden structure of data by reducing the high dimensional input space given by EXAFS spectral mixtures into a two-dimensional space. Thus, SOM enables an easier interpretation of the relationship between spectra and geochemical parameters such as the pH, concentration and temperature of samples. The fusion approach was generalised, as proposed by [5], to be able to combine more than two SOM (i.e., to implement several geochemical parameters simultaneously). For this, seven rectangular-shaped SOM were combined with 900 nodes per SOM (i.e., neurons). The first SOM (X-map) contains the EXAFS spectra, the second and following SOM (Y-maps) contains, for each spectrum in the X-map, the fractions

Fig. 103: Schematic presentation of three fused self-organising maps

(SOM) (X-, Y1-, and Y2-maps). The node which best matches a

randomly chosen prototype vector is determined.