Decadal prediction of terrestrial water storage and sea-level change using physics-informed neural networks

Terrestrial Water Storage (TWS) represents all sources of water on land, including polar ice sheets (Greenland and Antarctica), mountain glaciers, and hydrological components. Understanding the spatio-temporal evolution of TWS and the sea-level change remains a challenge in climate science, but is important in a variety of domains, including water resource management and sea-level rise in coastal areas. Machine learning has recently emerged as a viable technique for analyzing and predicting TWS. However, these models often lack interpretability, ignore uncertainty quantification, and are useful mainly for short-term prediction horizons of up to 1 year. Our study sets itself apart from previous attempts by presenting a novel probabilistic Physics-Informed Neural Networks (PINNs) methodology to (i) incorporate the physics of continental-ocean mass redistribution, (ii) quantify prediction uncertainty, and (iii) extend the prediction horizon of TWS to decadal (10-year) timescales. We present numerical experiments in the range 1900-2024 to show that our predictions are consistent with satellite gravimetry observations and improve on the projections of climate models by up to 42%, particularly for mountain glaciers. Furthermore, the predicted sea-level change agrees with the independent observations by satellite altimetry during 2003-2024, both in terms of fluctuations and the long-term trend within 0.1 mm/yr. These findings can advance our understanding of the spatio-temporal evolution of TWS and the corresponding sea-level change, both of which are of significant societal relevance.