Machine learning estimates of oceanic anthropogenic carbon (Cant)

“Machine learning, like all technology, does not always make the world a better place, but it can” (A. Moltzau, 2019 [20]).

Background

Since the beginning of the industrial revolution (1860), anthropogenic activities have increased the atmospheric carbon dioxide (CO2) from 280 ppm to over 400 ppm [6]. The global ocean has sequestered roughly a third of this anthropogenic CO2 (Cant, Fig.1), limiting the impacts on the global climate [3] such as the increase in the Earth’s mean temperature. However, Cant can only be estimated indirectly in the ocean with an uncertainty of ±20%. This uncertainty reduces our understanding of the processes associated with the Cant cycle and the reliability of future predictions.

Fig. 1: Global ocean anthropogenic carbon (Cant) column inventory. Data taken from the GLODAPv2 climatology [4], where they were estimated using the Transit-Time Distribution (TTD) method based on dichlorodifluoromethane (CFC-12) observations [7].

Problem

Cant estimates cannot be measured directly in the ocean being instead approximated using correlations with measurable variables (e.g. inorganic nutrients).

(1) The most commonly adopted methods for Cant estimates [2, 7] use Redfield ratios to isolate Cant from the ocean total dissolved inorganic carbon (tCO2) measurements. Redfield ratios are ratios of inorganic nutrients used to infer biological activities in the ocean [5]. Redfield ratios vary over space and time, based on the in-situ species.