Exploring the Complexities of Dissolved Organic Matter Photochemistry from the Molecular Level by Using Machine Learning Approaches.

Abstract

Dissolved organic matter (DOM) sustains a substantial part of the organic matter transported seaward, where photochemical reactions significantly affect its transformation and fate. The irradiation experiments can provide valuable information on the photochemical reactivity (photolabile, photoresistant, and photoproduct) of molecules. However, the inconsistency of the fate of irradiated molecules among different experiments curtailed our understanding of the roles the photochemical reactions have played, which cannot be properly addressed by traditional approaches. Here, we conducted irradiation experiments for samples from two large estuaries in China. Molecules that occurred in irradiation experiments were characterized by the Fourier transform ion cyclotron resonance mass spectrometry and assigned probabilistic labels to define their photochemical reactivity. These molecules with probabilistic labels were used to construct a learning database for establishing a suitable machine learning (ML) model. We further applied our well-trained ML model to "un-matched" (i.e., not detected in our irradiation experiments) molecules from five estuaries worldwide, to predict their photochemical reactivity. Results showed that numerous molecules with strong photolability can be captured solely by the ML model. Moreover, comparing DOM photochemical reactivity in five estuaries revealed that the riverine DOM chemistry largely determines their subsequent photochemical transformation. We offer an expandable and renewable approach based on ML to compatibly integrate existing irradiation experiments and shed insight into DOM transformation and degradation processes.