The Arctic Ocean covers 3 % of the Earth's surface but is estimated to contribute 5–14 % to the global ocean carbon sink. Sparse and unevenly distributed observations complicate our understanding of the size and the controlling mechanisms of this carbon sink. We adopt and advance the two-step neural network approach of Landschützer et al. (2016, https://doi.org/10.1002/2015gb005359; Self Organizing Map—Feed Forward Network) to improve region-specific reconstructions of the surface ocean partial pressure of carbon dioxide ((Formula presented.) ((Formula presented.))) in the Arctic Ocean and subsequently estimate the air-sea (Formula presented.) flux. Our study shows that biogeochemical properties previously selected as predictor variables for (Formula presented.) ((Formula presented.)) at the global scale are not well suited to the Arctic Ocean and a sensitivity study reveals large differences in the size of the Arctic Ocean carbon sink depending on the choice of air-sea (Formula presented.) flux parameterization. This is most acute for those relating to sea ice cover, leading to a difference of up to 25 % (9.2–13.3 Pg C) in the total size of the Arctic Ocean carbon sink over the 32-year duration of the study.