Depletion of firn air can lead to meltwater ponding. If conditions are unfavourable, it can cause hydrofracturing and disintegration of Antarctic ice shelves. Projections of firn air content can therefore be used in ice-sheet model simulations to prescribe which ice-shelf sections become vulnerable to collapse, and when. However, using a full firn model for simulating firn air content for a wide range of climate forcings is computationally expensive. We therefore developed an XGBoost emulator for firn air content, a fast machine-learning model trained to reproduce the firn air content from a full firn densification model. The emulator needs annual means of snowfall, surface melt, rain, air temperature, and wind speed as input. Using scenario and spatial blocking cross-evaluation, we find that the emulator predicts > 94 % of the variance in the firn air content of the full firn densification model. When a simulated firn air content of less than 2 m is assumed to represent the onset of melt ponding, the emulator shows 6–7 % false positives and 14–17 % false negatives. We demonstrate that surface melt is an important input variable, by constructing an alternative emulator that does not require surface melt. This emulator explains > 82 % of the variance in firn air content, and has poorer agreement with firn-model simulated melt pond locations. Including surface melt results in a better emulator, because the non-linear behaviour of snow surface albedo is hard to predict from atmospheric variables only.