Forecasting the localized bilateral effects of ocean acidification on the counter carbonate pump using recurrent neural networks

This is a Preprint and has not been peer reviewed. This is version 6 of this Preprint.

Add a Comment

You must log in to post a comment.


Comments

There are no comments or no comments have been made public for this article.

Downloads

Download Preprint

Supplementary Files
Authors

Eshan Ramesh 

Abstract

The counter carbonate pump(CCP) is responsible for carbon dioxide sequestration and cycling forms of carbon in the ocean. It is primarily driven by calcifying plankton, such as foraminifera, coccolithophores, and pteropods. These organisms are particularly vulnerable to ocean acidification, which can have disastrous effects on their skeletons and productivity, upsetting the marine carbon cycle in ways that have not been quantified due to their chaotic nature. This project aims to provides high-resolution, accurate, and robust predictions of the efficiency of the CCP. These predictions are achieved by training recurrent neural networks on satellite-determined particulate inorganic carbon(PIC) to particulate organic carbon(POC) ratio data that represents the efficency of calcifying organisms in exporting PIC. Recurrent neural networks mimic the way in which biological neurons learn patterns and recall past experiences, and have been shown to be good at predicting chaotic time series. For each 9x9km square sample in a random subset of the dataset, the time series of historical PIC, POC, tropospheric satellite CO2 data, latitude and longitude were used by the model. The applications of the model are threefold. First, it can pinpoint the most vulnerable area in the ocean at arbitrary times in the future. Second, general trends in PIC are a proxy for ocean acidification. Finally, anomalous spikes in PIC can be potential coccolithophore blooms, which are dangerous for sub-surface marine life. The model predicts the ratio value to an error of 1.3\%, tested using cross validation, and is significantly better than linear regression. Future directions for work include physically testing affected specimens in projected water characteristics.

DOI

https://doi.org/10.31223/osf.io/cyq3e

Subjects

Artificial Intelligence and Robotics, Chemistry, Computer Sciences, Environmental Chemistry, Numerical Analysis and Scientific Computing, Oceanography, Oceanography and Atmospheric Sciences and Meteorology, Physical Sciences and Mathematics

Keywords

machine learning, Coccolithophore, Neural Networks, carbonate cycle, ocean acidification

Dates

Published: 2020-07-08 12:44

Last Updated: 2020-10-24 17:23

Older Versions
License

GNU Lesser General Public License (LGPL) 2.1