Saccharide Concentration Prediction from Proxy Ocean Samples Analyzed Via Infrared Spectroscopy and Quantitative Machine Learning

Abstract

Solvated organics in the ocean are present in relatively small concentrations but contribute largely to ocean chemical diversity and complexity. Existing in the ocean as dissolved organic carbon (DOC) and enriched within the sea surface microlayer (SSML), these compounds have large impacts on atmospheric chemistry through their contributions to cloud nucleation, ice formation, and other climatological processes. The ability to quantify the concentrations of organics in ocean samples is critical to understanding these marine processes. The work presented herein details an investigation to develop a machine learning (ML) methodology utilizing infrared spectroscopy data to accurately estimate saccharide concentrations in complex solutions. We evaluated multivariate linear regression (MLR), K-nearest neighbors (KNN), decision trees (DT), gradient-boosted regressors (GBR), multilayer perceptrons (MLP), and support vector regressors (SVR) toward this goal. SVR models are shown to best predict accurate generalized saccharide concentrations. Our work presents an application combining fast spectroscopic techniques with ML to analyze organic composition in proxy ocean samples. As a result, we target a generalized method for analyzing field marine samples more efficiently without sacrificing accuracy or precision.