The basal thermal state of the Antarctic ice sheet strongly influences ice dynamics and mass balance. Although basal thermal states can be simulated using ice-sheet models, significant uncertainties persist, with poorly constrained geothermal heat flux being one important source of uncertainty. The substantial computational cost of ice-sheet models restricts large ensemble simulations essential for uncertainty quantification and sensitivity analyses. To address this limitation, we develop machine-learning emulators for rapidly simulating the basal temperature and melt rate of Totten Glacier. Three distinct machine-learning emulators were trained using outputs of full-Stokes model simulations with different geothermal heat flux fields, achieving accurate replication ( R 2 up to 0.92). At least five simulations are needed to train a reliable emulator for basal ice temperature. Model interpretability reveals complementary physics learning: Random forest and XGBoost prioritize local dynamic drivers (basal topography, ice velocity), while neural network captures lower-complexity statistical mapping that is more strongly influenced by globally predictive variables (surface temperature, ice thickness). For basal melt rate, surface velocity emerges as the primary driver. These findings highlight substantial potential for integrating machine learning with ice-sheet modeling to efficiently and better predict basal thermal states.