38-1 Scaling up Speciation: Using Machine Learning to Upscale Speciation Measurements in Arsenic-Impacted Groundwater

Session: Geoscience and Hydrogeology in the AI Era: From Predictive Models to Real-Time Applications

Presenting Author:

Authors:

Abstract:

Naturally-occurring contamination of groundwater with arsenic (As) affects millions of people worldwide and leads to adverse health effects, including various cancers. Arsenic release to groundwater is driven by biogeochemical redox processes: typically attributed to the reductive dissolution of As-bearing iron (III) minerals in the anoxic subsurface. However, many factors influence As fate and transport in groundwater, making it a long-standing problem despite decades of research.

Importantly, understanding speciation has emerged as critical — both in the solid and aqueous phase. For example, in the solid-phase, iron (Fe) minerals have long been known to be the source of As into groundwater. After the reductive dissolution of As-bearing Fe (III) minerals, subsequent Fe mineral transformations into more reduced Fe minerals affects remaining sorption site capacity for As. Indeed, in groundwater As hotspots, finding reduced Fe(II) minerals in aquifer material is indicative of extensive reduction that has likely released considerable As into groundwater. Meanwhile, measurements of aqueous speciation have been far less extensive than that of solid-phase speciation. For example, usually only total concentrations of As are measured in groundwater. However, in recent years, it was discovered that certain aqueous species of As, such as thiolated As species (e.g., aqueous As species with at least one sulfide group attached instead of the usual As oxyanion) sorb poorly onto most Fe minerals. Thiolated As species can thus move more quickly in the subsurface and lead to faster contamination. Yet, despite its importance, speciation is not usually measured beyond a specific field site and is practically impossible to measure on a large scale – either in the solid or aqueous phase.

Here, we demonstrate that using machine learning is extremely valuable for upscaling datasets of speciation, whether in the solid or aqueous phase. We use statistical relationships between typically characterized measurements (such as total concentrations) to upscale and derive novel insights from speciation measurements: with case studies ranging from upscaling synchrotron X-ray absorption spectroscopy measurements of aquifer sediment cores to predicting and verifying aqueous As speciation measurements with random forest modeling in Vietnam. By using machine learning, information gained from precious speciation measurements — due to time-limited synchrotron-based analyses or tedious preservation of aqueous speciation in remote environments — is critically upscaled to generalize insights into this complex hydrogeological problem.

Geological Society of America Abstracts with Program. Vol. 57, No. 6, 2025

doi: 10.1130/abs/2025AM-8257

© Copyright 2025 The Geological Society of America (GSA), all rights reserved.