Global Estimates of Biogenic Methane Production in Marine Sediments Using Machine Learning and Deterministic Modeling

We have developed a model of geospatially estimating carbon accumulation and methanogenesis in seabed sediments that uses more accurate and sophisticated inputs to models than used in previous estimates. Using this hybrid stochastic and deterministic model, we estimate the maximum carbon available for methanogenesis in the global seabed, and subsequent microbial methane generated as a function of location and depth (including the gas hydrate stability zone). Global integration over present and previously microbially reactive sediments column yields total carbon and methane to be similar to 0.8-2.2 x 10(6) and 1.1-3.0 x 10(6) Pg C and CH4, respectively. Our improvements to accuracy include using geospatially machine learned estimates of seafloor inputs to which the methanogenesis modeling is most sensitive (e.g., total organic carbon, heat flux, porosity). Our improvements to model sophistication include geospatially dependent modeling (on a 5 x 5 arc-minute grid), a new model of sediment compaction (allowing for non-linear geothermal gradients), and variable age versus depth at each grid cell. A carbon reservoir of the magnitude we estimate here is consistent with the recent IPCC suggestion that long-term carbon sinks could explain imbalances in reduction of atmospheric CO2 over the last 50 million years. Our technique provides a foundation of using globally updateable machine learning parameters as the input to geologic and geochemical models, allowing for new observations to update global budgets of carbon available for methane, and subsequent total estimates of seabed methane.