Artifical neural networks - another view on actinide chemistry

Driven by the enormous increase of computing power in the last few years, several theoretical models for biological neural networks have been developed and new modelling concepts are coming up. Close to their biological counterpart, artificial neural networks (ANNs) have been developed, which can be trained by presenting input data along with the desired output [1]. ANNs are used for example to recognize hand-written text, human faces, natural speech or personal interests, all features now commonly used by web-based social networks like Facebook or by search engines like Google. Another step forward was the development of spiking neural networks, a third generation ANN, which counts at the moment as the closest approximation of biological neural networks [2] and is strongly discussed in the community [3]. In general, ANNs can be considered as a method to infer functional relationships between a cause and the resulting effect. Note that these functional relationships are not inferred from mathematical models; first, the mathematical treatment would be too complex, and second, the required mathematical simplifications by using justified assumptions and constraints would most probably lead to a poor description of the problem. Therefore ANNs are seen as a more efficient and more accurate method to describe complex functional relationships. Here we demonstrate using two examples, that selforganizing maps (SOM) [4,5], a special kind of ANN, are well suited for analysing the complex chemistry of actinides.
The first example shows the relationships between the structure of aliphatic ((di-)hydroxy-)carboxylic acids and their complexation mode towards uranyl (UVI). For this, U LIII-edge EXAFS spectra from 13 aliphatic carboxylic acids (acetic, succinic, tartaric, lactic, 3-hydroxybutyric, citric, formic, malic, maleic, malonic, oxalic, propionic, and tricarballylic acid) at a range of pH, uranium and ligand concentrations were measured, resulting in 60 EXAFS spectra [6]. Based on the known structures of the ligands, SOM was used to determine the dependencies between the structure of the UVI carboxylate complexes and the structure of the interacting ligands, and to derive a predictive classification of the former. SOM revealed, for instance, that acids with an OH-group in α-position cause the formation of monomeric chelates and dimeric and trimeric UVI complexes, while an OH-group in β-position leads only to monomers, where uranyl is bidentately coordinated to the carboxylic group. In the second example, we apply SOM to the U LIII-edge EXAFS spectra of UVI sorption complexes with Al(hydr)oxides in order to determine the dependency of the structure of the formed sorption complexes on relevant physicochemical parameters like pH, pCO2, surface area, and surface loading. SOM here clearly reveals, for example, that polynuclear sorption complexes become predominant the higher the pH and the surface loading, while at low pH mononuclear complexes are present either as binary complexes or – in the presence of carbonate - as ternary complexes.
A properly trained SOM can also be used for the prediction of the spectra and the fractions of the complexes for a given ligand and a set of physicochemical parameters, hence SOM can replace thermodynamic speciation calculations, in case complex formation constants are not available. In turn, for a given spectrum, the corresponding physicochemical parameter can be predicted. SOM is not only restricted to EXAFS, but can use input from any other spectroscopy like NMR, UV-vis, infrared and Raman, or from diffraction/scattering patterns -together with chemical information- in order to derive a reliable multiscale speciation of actinides.