Predicting radioactive waste glass properties with machine learning and modelling
Over several decades, numerous glass dissolution and characterisation experiments have been performed by the nuclear industry, which currently vitrifies its high-level radioactive waste (HLW), ultimately in preparation for deep geological disposal. This data has largely been collected in isolation, and there has been an open question whether this data could be of value when used collectively. With this motivation, this PhD has aimed to either predict unstructured glass property data using machine learning or apply mechanistic modelling on a large scale both to evaluate the impact of glass within the wider geological disposal setting and understand glass precipitate formation.
Machine learning was first applied to predict static leaching, specifically boron releases, having amalgamated UK vitrification campaign, literature, and unpublished laboratory data. Accuracy was shown to be high having partitioned training/test data on a whole-experiment basis, on individual time-point data, and when forecasting, provided that features include pH or release data, and typically a tree-based algorithm, for example, a bagged random forest. The trained algorithms could successfully predict (> 12 year) long-term UK MW glass and newly measured (7 day) short-term ISG/MW25 free pH and (27 day) ISG/Ca-Zn Tris-Tris HCl buffered pH dissolution. The short-term free pH experiments complemented longer-term leaching experiments, and TEM imaging of 90 °C MW25 samples revealed significant alteration, even after 1 day of leaching. The buffered pH experiments found Ca/Zn release rates to be greater than that of ISG, with SEM/EDS analysis suggesting that the high alteration rate is likely a result of the formation of a macro-porous (likely phyllosilicate) layer that was increasingly developed with increased Mg oxide in the pristine Ca/Zn glasses.
Machine learning was secondly applied to predict B log/non-log initial dissolution rates (IDR) in complex glasses. The use of large datasets obtained from a variety of sources, covering a diverse range of glass compositions, shows an accurate performance that is comparable with similar methods applied to simplified non-nuclear glasses from more limited datasets. Machine learning could also predict Si, Na, and Al initial dissolution rates in simplified Na-Al-Si glasses, with predictive performance being higher if replacing final pH with a species (Si, Na, or Al) dissolution rate as a learning feature, although there is no preferred output species (Si, Na, or Al), despite the solubility difference between these species.
Machine learning algorithms were subsequently trained on an international database (ALTGLASS) and shown to accurately predict ALTGLASS test data (B, Li, Na, and Si releases), depending on the machine learning algorithms and features used. For example, composition and experimental setup variables alone were insufficient, with experimental release needing to be included for accurate learning. Trained algorithms were subsequently applied to independent data, with release errors typically being overestimated for B, underestimated for Li, becoming more substantial with increasing Na releases, and primarily overestimated for Si. Machine learning was further successfully applied to various glass viscosity, viscosity model fitting, density, and phase separation nuclear industry data.
Glass dissolution was then analysed in the larger geological disposal facility setting by using Monte-Carlo reactive-transport modelling, building on the work of Iwalewa and Farnan¹ to examine radionuclide migration near to the disposal vault (‘crown’). High-level radioactive waste (HLW) glass base-line simulations indicated that there is no preference in having granite versus clay host rocks for a given canister type with respect to minimising activity near to the disposal vault but that a copper canister is preferable to steel. Intermediate-level radioactive waste (ILW) glass simulations suggested that a granite-bentonite-cast-iron environment yields lowest crown activities with cast-iron preferable to concrete as the canister, bentonite preferable to cement as the buffer/backfill, and granite preferable to clay as the host rock. Varying glass dissolution source terms (initial, residual and resumption rates) had a logical but small effect on radionuclide migration compared with the HLW and ILW base-line models. Moreover, random variation on the overall glass degradation rate gave a peak crown activity that was negligibly different from the HLW (or ILW) base-lines.
Speciation simulations systematically performed on increasingly complex chemical systems (Si, Si-B, Si-B-Na,...Si-B-Na-...-C) highlighted the precipitates that may thermochemically form in various glass systems. A limited number of precipitates were identified (for Al, Si, Mg, and Ca species, etc.) from speciation simulations having used various experimental leaching data in the model input parameters. Differences in the precipitates observed may guide the modelling of glass leaching experiments. Finally, analytical/numerical glass reactivity with allowance for the alteration layer (GRAAL) model optimisation identified the conflict in fitting different leached species solution concentration time behaviour.