Abstract

Marine sediments represent the most important sink for natural organic matter (NOM) across geological time spans, in which carbon-containing molecules are sequestered away and can escape remineralization to CO2 by microbial degradation. Strong associations between iron oxide minerals and organic matter reaching the seafloor play a fundamental role in this preservation and have been known for some decades. Despite the importance of this protective mechanism in the balances of the global carbon budget, very little is known about the affinity of NOM for reduced iron species such as mackinawite (FeS) in the anoxic layers of sediment. In this study, equilibrium partition coefficients (Kd) for three types of NOM (soil leachate, corn leaf and plankton lysate) on FeS were determined through batch sorption experiments. Attenuated Total Reflectance Fourier Transform Infrared spectra (ATR-FTIR) of the organic matter was then used as model inputs to train a partial least squares regression model (PLSR) to quantitatively predict Kd values based on the FTIR spectra of NOM. The final model fit the training data with an R2 of 0.97 (n = 17) and the validation data with a Q2 of 0.98 (n = 5) and a RMSEP of 0.036. Inspection of the PLSR regression coefficients indicate that functional groups characteristic of polysaccharides are the greatest positive predictors of NOM sorption onto FeS at sediment pore water pH. To our knowledge, this research presents a novel machine-learning approach to the quantitative modelling of NOM sorption to minerals found in marine environments.