Sub-grid-scale processes occurring at or near the surface of an ice sheet have a potentially large impact on local and integrated net accumulation of snow via redistribution and sublimation. Given observational complexity, they are either ignored or parameterized over large-length scales. Here, we train random forest models to predict 1-km variability in net accumulation over the Antarctic Ice Sheet using atmospheric variables and topographic characteristics as predictors. Observations of net snow accumulation from both in situ and airborne radar data provide the input observable targets needed to train the random forest models. We find that kilometer-scale processes modify local net accumulation by as much as 172% of the atmospheric model mean. The correlation in space between the predicted net accumulation variability and satellite-derived surface-height change indicates that kilometer-scale processes operate differently through time, driven largely by the seasonal anomalies in snow accumulation.