ALBERT

Description: Sahel summer precipitation is highly variable with floods and droughts occurring on a regular basis. Summer rainfall events were predicted using neural networks based on a collection of preseason climate indices. Understanding the prediction of a neural network is limited as this model type is a black-box. Layer-wise relevance propagation was applied to account for this issue. It allowed the attribution of single predictions to certain states in the underlying data. By that the trustworthiness of a neural network can be assessed and driving mechanisms of climate events can be pointed out. While the presented rainfall predictions are worse than earlier attempts from other studies, and consequently any event attribution is not reliable, this study demonstrates how layer-wise relevance can be implemented in climate-related contexts and illustrates the potetial of that technique in this field.

Description: While atmospheric CO2 concentrations have been increasing during recent decades due to anthropogenic emissions, the ocean has acted as a sink for atmospheric carbon. Essentially, the global air-sea flux of CO2 showed a trend towards more oceanic uptake as expected from increasing emissions. Yet, the oceanic CO2 uptake also responded to climate change and fluctuated due to climate variability and variations in the growth rate of atmospheric CO2. So far, the drivers of the variability in oceanic CO2 uptake are not conclusively understood. In this thesis, the global ocean biogeochemistry model FESOM-1.4-REcoM is used to quantify the effects of climate change and of the increasing atmospheric CO2 concentration on the trend in the oceanic carbon uptake during the period 1958-2019 (62 years). Two approaches are applied: (1) Offline diagnostics based on a linear approximation relating the trends in the sea surface temperature, dissolved inorganic carbon, alkalinity, salinity plus freshwater fluxes, wind velocity and sea-ice concentration to the trend in the CO2 flux and (2) a model experiment with the historical forcing fields compared to simulations in which certain forcing fields (e.g. winds and the atmospheric forcing fields that control the sea surface temperature) are replaced by a repeated year forcing in order to isolate their effects on the CO2 flux. In FESOM-1.4-REcoM, the ocean took up 1:85 PgCyr-1 of atmospheric CO2 on average during the simulated period. The ocean carbon sink increased with a trend of 23:8 TgCyr-1 per yr. In a simulation with rising atmospheric CO2 concentrations but without climate change and variability, the trend in oceanic carbon uptake was 27% higher than that, suggesting that climate variability has substantially reduced the uptake over the simulated period. Of this, a trend towards more outgassing of 2:9 TgCyr-1 per yr was driven by the change and variability in winds, which was particularly relevant in the polar and subpolar regions. Hereby, a comparison between the offline and online approach reveals that the effect of winds was dominated by wind-driven changes in the transport of natural carbon with the circulation. Global warming caused a trend towards more oceanic outgassing of 2:3 TgCyr-1 per yr, which mostly originated from the tropical and subtropical zone. The increasing sea surface temperature led to more outgassing due to the reduced solubility of CO2. The offine estimate for the effect of warming on the trend in the CO2 flux is much larger, which can be attributed to the neglect of compensating feedbacks. In particular, the simulated effect of global warming reveals that in response to the increasing temperature, the concentration of dissolved inorganic carbon in the mixed layer decreased, which attenuated the thermally-driven outgassing. Changes in all other variables were less important drivers of the trend in the CO2 flux.