ML for Surface Complexation Model Development

Abstract:

Surface complexation models (SCMs) describe adsorption-desorption reactions of heavy metals and radionuclides in the geological environment, which plays a critical role in many environmental science applications such as contamination risk assessments and nuclear waste disposal. Datasets from laboratory measurements are used to build SCMs as well as to determine key parameters such as equilibrium constants. The datasets however, are collected by multiple institutions, which are building specific models. There is no framework to compile and analyze all the datasets available and to test different models across common datasets. In addition, SCMs include a significant number of reactions and parameters which leads to overfitting, over-parameterization and or significant computational requirements when coupled to contaminant transport models. Recently, there has been an effort to compile all the experimental datasets across the world, which offer a unique opportunity for improving SCMs coupled with recent advances in machine learning. In this work, we develop a suite of machine learning algorithms to support an effective SCM development. We establish a data analytics workflow of sorption experiment data, including identifying correlations among multiple variables and key controls. In particular, we apply principal component analysis and clustering methods to identify the bias associated with institutions and solid material. In addition, we develop a framework to compare multiple SCMs against common datasets, by including the Bayesian information criteria and other information theoretic metrics. In parallel, we explore an alternative way to parameterize aqueous-solid partitioning coefficients (Kd) dependent on pH, redox, and other groundwater conditions. We compare different machine learning-based methods such as Random Forest. With this data-driven Kd – a smart Kd – we can reduce the complexity of SCMs and to include variable sorption behaviors in contaminant transport models while avoiding computationally expensive geochemical simulations.