A Machine Learning Approach to Constrain Ocean Warming in the Atlantic since 1945.

A machine learning approach to interpolating ocean temperature data is applied in the Atlantic Ocean to produce sub-annual estimates of ocean heat content for the period 1945-present. Using in-situ data from the well sampled period between 2006-present (the Argo Era) and running iteratively backwards, a time-series prediction model capable of filling gaps in ocean temperature data is constructed for each of the 102 standard depth levels (from 0-5500 m) in the World Ocean Atlas. This method is validated by adding realistic noise to ocean temperature data in 2 CMIP 6 climate model runs that have been reduced to observational sparsity. The gaps in the climate models are refilled and compared to the originals to validate this machine learning approach. When applied to available observations this method produces fields of temperature anomalies useful for constraining spatio-temporal trends of ocean warming, including assessing ocean heat content for the abyssal Atlantic Ocean (depths greater than 2000 m).