Performance assessment models that inform society on the long-term risks associated with nuclear waste storage hinge upon the underpinning thermodynamic descriptions of retarding subsurface mineral-radionuclide reactions. To this extent, a wealth of surface complexation models (SCMs) has been developed over the past many decades to be incorporated into these large-scale analyses. However, the advancement of SCM development has slowed due to the many rigid assumptions posed by these models. Divergent assumptions on electrostatic treatments of surface charge accumulation and the use of different thermodynamic constants and databases leave the scientific community conflicted and unable to make clear comparisons between these various historic models. However, recent advances in machine learning (ML) approaches provide a promising new way to consolidate geochemical data associated with mineral-radionuclide reactions, outputting a more holistic, universal result for metal-mineral stability. In this presentation, we discuss quartz-uranium surface complexation as a case study to exploit recent data digitization efforts and the powerful application of random-forest ML algorithms to predict the partitioning of uranium onto quartz surfaces from aqueous solutions. While ML approaches have been met with some skepticism in the sorption science community, we attempt to bridge the gap between novel ML and traditional SCM methods through the development of ‘investigatory’ features that are inputted into the random-forest ML process. We demonstrate improvement in model U(VI) sorption prediction capabilities (R<sup>2</sup> = 0.90 to R<sup>2</sup> &gt; 0.97) through incorporating features derived from traditional SCMs, such as concentrations of relevant, dominant aqueous complexes, the incorporation of gas-liquid exchange relationships for carbon dioxide, and an explicit variable for adsorbate/adsorbent ratios (also known as surface excess). While ML approaches have historically been associated with unsupervised black-box models that oversimplify complex geochemical relationships, we readily incorporate decades’ worth of SCM knowledge by providing the random-forest learning algorithm with theoretically derived, mechanistic pieces of sorption information. Ultimately, the authors of this work present a new method to ameliorate past skepticism of ML methods and provide a synthesis of the old and the new: the application of ML sorption models informed by traditional SCM principles.

Unifying machine learning and traditional surface complexation modeling approaches to shine new light onto radionuclide-mineral interaction predictions

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.