Arctic sea ice has steadily declined under global warming, whereas Antarctic sea ice, after a modest multi-decadal increase, began a sharp downturn in 2016 and reached a record minimum in February 2023. These recent changes highlight the need for improved seasonal prediction of Antarctic sea-ice variability, particularly during austral summer when conventional numerical models often have limited skill. Here, we develop a deep-learning framework for seasonal prediction of Antarctic sea ice using a selective set of physically relevant atmospheric and oceanic predictors. The model substantially improves predictability in austral summer while maintaining a physically interpretable representation of the underlying drivers. To go beyond prediction skill alone, we examine the model using spatially explicit and seasonally resolved explainable artificial intelligence contribution maps. These reveal distinct seasonal predictor influences: meridional wind (V) dominates in austral summer (Dec–Feb), whereas downward shortwave radiation (SW) is most important in austral spring (Sep–Nov). The influence of V on sea ice differs between summer and winter, while springtime SW contributions show marked regional contrasts, suggesting that the model captures cloud-mediated feedbacks. Together, these results demonstrate that deep learning can not only improve seasonal Antarctic sea-ice forecasts but also provide physically interpretable insights into atmosphere–ocean–sea-ice interactions across space and time.