MAPPING SILICATE WEATHERING FLUX WITH COUPLED MACHINE LEARNING TECHNIQUES

The chemical weathering of silicate rocks in continental settings shapes landscapes, controls soil evolution, and supplies nutrients to the global biosphere. During weathering process, the reaction of silicate rocks with CO₂ also acts as a net carbon sink and provides negative feedback in the global carbon cycle. As such, silicate weathering is thought to have played a key role in keeping surface environments clement throughout the majority of Earth’s history. Despite continual advances in simulating silicate weathering fluxes, traditional modeling approaches still have limited power to reproduce empirical estimates. As such, a quantitative framework that yields robust prediction of silicate weathering rates both in local and global scales remains a challenge. In this study, we apply two data-driven machine learning (ML) techniques to map the silicate weathering flux over the Contiguous United States (CONUS). Specifically, we collect and harmonize available gauge station observations and long-term environmental and geological datasets in corresponding watersheds over the CONUS land surface. We then apply non-negative matrix factorization (NMF) ML model to apportion the river bicarbonate to weathering of silicates or carbonates, which will yield the silicate weathering flux for each watershed. This NMF step is critical because it boosters the number of watersheds that could be used in the following ML model. The silicate weathering fluxes and corresponding watershed properties are then fed into the second ML model based on tree ensembles to build the final ML framework, which is subsequently used to map the silicate weathering flux over the CONUS land surface. Besides, the relative importance of key parameters (such as the climatic/hydrological conditions, crustal compositions, and tectonics) in regulating silicate weathering rates are delineated from the data-driven perspective.