Predicting Geothermal Heat Flow in Antarctica With a Machine Learning Approach

Abstract

In-situ observations from temperature gradient measurements are sparse and go along with several uncertainties like climatic changes or hydrothermal circulation (Burton-Johnson et al., 2020).To establish continent-wide heat flow models, one must refer to indirect methods using geophysical or geological data. The results differ immensely, for example, between magnetic and seismological data (e.g., An et al., 2015b;Martos et al., 2017), and the underlying assumptions cannot be easily combined (Lösing et al., 2020). While these different approaches usually take the respective sensitivity ranges into account, simplifications like a definition of laterally constant thermal parameters are not considered in the assessment. Recently, Shen et al. (2020) demonstrated how an updated seismic tomography model completely changed the estimated GHF and its uncertainties compared to earlier studies by Shapiro and Ritzwoller (2004). In ice-covered regions, different geophysical models show no consensus on magnitude and spatial distribution of heat flow (Rezvanbehbahani et al., 2019;Van Liefferinge, 2018).

Greenland Geothermal Heat Flow Database and Map (Version 1)

“…This approach was initially presented for Greenland by Rezvanbehbahani et al (2017) and subsequently revised for Antarctica by Lösing and Ebbing (2021) (https://github.com/MareenLoesing/GHF-Antarctica-MachineLearning). Lösing and Ebbing (2021) enhanced the machine learning algorithm by using an advanced, and more regularized, gradient boosting regression and provided more detailed evaluation of the influence of regional and global geophysical datasets. This evaluation showed the added value of applying well-constrained regional data, as global datasets often have a high uncertainty in polar regions.…”

“…This evaluation showed the added value of applying well-constrained regional data, as global datasets often have a high uncertainty in polar regions. More technical descriptions on the method can be found in Rezvanbehbahani et al (2017) and Lösing and Ebbing (2021) and a graphical overview of these datasets is provided in the Appendix of this study.…”

“…Following Lösing and Ebbing (2021), we optimize a global supervised machine learning regression approach by incorporating regional datasets best suited for Greenland (Table 5). We combine the heat flow measurements described above with the global point dataset used by Lösing and Ebbing (2021).…”

“…Following Lösing and Ebbing (2021), we optimize a global supervised machine learning regression approach by incorporating regional datasets best suited for Greenland (Table 5). We combine the heat flow measurements described above with the global point dataset used by Lösing and Ebbing (2021). All the global or regional predictive geophysical datasets are similarly interpolated from their native resolution to a common 55 km grid, which defines the fundamental resolution of the final heat flow solution.…”

Review of Colgan and others, "Greenland Geothermal Heat Flow Database and Map (Version 1)" [essd-2021-290]

“…The importance parameter evaluates the relative importance of each input dataset for predicting the results. More details about the calculation and theory of the importance parameter can be found in Lösing and Ebbing (2021). For the continental model domain, the distance to volcanoes is the most important feature, followed by the Moho depth and the tectonic regionalization.…”