Can neural networks predict ice sheet collapse decades ahead?

Long-range forecasts of ice sheets fail because errors pile up when you chain one-step predictions together over decades. This work trains a single graph neural network to predict multiple future time steps at once—learning direct state-to-state transitions at 10, 20, or 50 years ahead rather than rolling out step-by-step. By predicting increments relative to the current state and using selective refinement during inference, the model stays stable and accurate far longer than autoregressive baselines on glacier simulations.