Mapping Antarctic Geothermal Heat Flow with Deep Neural Networks optimized by Particle Swarm Optimization Algorithm

Abstract. The spatial distribution of geothermal heat flow (GHF) beneath the Antarctic Ice Sheet is a major source of uncertainty in projections of ice sheet dynamics and sea-level rise. Direct measurements are sparse, necessitating robust modeling approaches. In this study, we developed a neural network framework whose architecture and hyperparameters are optimized using a particle swarm optimization (PSO) algorithm. Trained on a global heat flow compilation and a suite of geophysical datasets, our model generates a new GHF map for the entire continent. The model's accuracy in regions lacking direct measurements was confirmed through training density validation, with prediction errors constrained to within 20 %. The resulting map delineates a distinct dichotomy: East Antarctica exhibits predominantly low GHF values (<60 mW m^-2) with notable exceptions of high heat flow (>80 mW m^-2) in the Vostok Subglacial Highlands and Gamburtsev Subglacial Mountains. In contrast, West Antarctica is characterized by widespread high heat flow (>60 mW m^-2), especially in tectonically active regions like the Transantarctic Mountains and the Amundsen Sea sector. These predictions show agreement when compared with direct borehole measurements. Our work offers a new, robust estimate of Antarctic GHF, providing a critical boundary condition for ice sheet models. We suggest that future improvements in accuracy and interpretability can be gained by assimilating more high-resolution drilling data and integrating physical constraints into the model framework.