Mapping Antarctic geothermal heat flow with deep neural networks optimized by particle swarm optimization algorithm

Abstract. Geothermal heat flow (GHF) beneath the Antarctic Ice Sheet (AIS) is a critical basal boundary condition for ice-sheet dynamics modelling and sea-level rise projections. However, it remains insufficiently constrained due to the limited availability of in-situ observations. Here, we propose a deep neural network (DNN) framework that integrates both Particle Swarm Optimization (PSO) and a Bayesian output layer to predict GHF across the entire AIS. Rather than relying on manual or localized parameter selection, the PSO algorithm uses a robust global search mechanism to autonomously optimize key DNN hyperparameters, while the Bayesian layer provides probabilistic GHF predictions and rigorous uncertainty quantification. Model validation based on European GHF datasets demonstrates that the proposed framework consistently outperforms conventional approaches under data-sparse conditions. Then, applying the model to the AIS, we estimate that the Antarctic GHF ranges from 20 to 110 mW m−2 , with a continental mean value of 65.6 mW m−2 . Elevated GHF values (generally > 70 mW m−2 ) dominate much of West Antarctica, while localized high-anomaly zones are identified in parts of East Antarctica, including the Subglacial Lake Vostok region. Uncertainty mapping and the decomposition analysis reveal that most of the uncertainty is inherited from the underlying measurements, indicating that more high-quality observations are needed to further constrain Antarctic GHF predictions.