Predicting Langmuir model parameters for tungsten adsorption in heterogeneous soils using compositional signatures

Abstract

• Soil taxonomic designations were statistically discriminated by compositional data. • Stacked ensemble models showed greater accuracies than unstacked models. • Multiple W chemical interactions evident across diverse range of soil “types”. Metallic tungsten (W) is a highly dense material of increasing importance to the U.S. Army as a strategic, non-radioactive replacement for depleted uranium. While there is a growing body of evidence regarding the mechanistic behavior of ionic W (formed after the spontaneous oxidation of metal) in the environment, predicting its environmental fate remains challenging, owing to the widespread geochemical heterogeneity of soils. Therefore, we developed W adsorption prediction models by creating different functional “compositions” of the chemical and physical characteristics for different soil “types” (a non-specific yet commonly used to term to designate different soils). A relatively small dataset consisting of twenty soils (possessing six different soil “types” from across the U.S.) were evaluated for W adsorption behavior. Physical and chemical soil data were separated into water-extracted (WE), bulk, and particle-size distribution (PSD) compositions, and center log-ratio (clr) transformed. Classification models built using extremely randomized trees (ERT) showed that the compositions' accuracies were WE > Bulk > PSD at the Order and Suborder levels. W's adsorption isotherms were constructed using batch equilibrium experiments and modeled against the Langmuir model, where S max = calculated adsorption maximum, K 1/L = inverse Langmuir affinity coefficient. Afterward, both the ERT and ensemble, or stacked, ERT models (by addition of Order and/or Suborder taxonomic labels as ensemble classifiers) were developed for predicting the S max and K 1/L parameters based on the different compositions. In general, model accuracies were substantially increased by the addition of the labels (stacked models). Feature importance calculations pointed to a wide range of potential chemical mechanisms simultaneously controlling W adsorption, laying the groundwork for more detailed in-situ elemental speciation studies. Overall, this work showcased a new technological capability allowing for accurately predicting W adsorption on a wide variety of morphological soil designations. Capsule: This work found that soil morphological designations greatly improved the accuracy of Langmuir adsorption predictions of CoDA-transformed characterization data.