Deep neural network surrogate for surface complexation model of metal oxide/electrolyte interface

Abstract

HypothesisSurface Complexation Models (SCM) are one of the most successful geochemical thermodynamics models of the speciation and charging of the mineral/electrolyte interface. However, finding unique and transferable SCM parameter values by fitting experimental data using existing numerical solvers is often challenging. We hypothesized that SCM surrogates built using Artificial Intelligence algorithms provide an alternative strategy for determining model parameter values and thus offer an ultrafast prediction of electrical double layer characteristics of metal oxide/electrolyte interface. ExperimentsWe generated synthetic datasets of surface charge and electrokinetic potential as a function of pH, ionic strength, varying site densities, acidity, and ion-affinities of mineral surfaces using the 2-pK Triple Layer Model - one of the most used generic SCM. We explored a wide range of types of mineral surfaces using random walks in the parameter space. Next, we trained and validated two AI surrogates using our synthetic SCM data. FindingsWe showed that AI-SCM surrogates for oxide/electrolyte interfaces are possible and cost-effective. They allow ultrafast and accurate SCM parametrization using synthetic data. AI-SCM can predict thermodynamic parameters from just a few titration data points, such as ion affinities and proton binding constants. What is more, the computational strategy we presented here to surrogate geochemical models by pre-trained AI architectures is more universal and can be applied to other geochemical/thermodynamic models – replacing the computationally expensive solvers with trained AI, thus minimizing computing costs and speeding up data analysis.