Abstract
<jats:p>Understanding the causes and socioenvironmental implications of sea-level change is of fundamental importance in the field of geodesy, solid Earth geophysics, polar studies, and cryospheric sciences. Sea-level change is believed to be driven by three main processes: (i) barystatic: continental-ocean mass redistribution due to terrestrial water storage anomalies, (ii) steric: ocean thermal expansion (thermosteric) and salinity variations (halosteric), and (iii) glacial isostatic adjustment: solid Earth rebound due to viscous mantle relaxation (we remove the effect of the more sporadic process of vertical land motion). Current studies typically focus on one of the aspects (i)-(iii) and the total sea-level change (the sum of all three components) is derived by adding the disparate estimates of various studies. This approach introduces substantial uncertainties and errors in sea-level estimates. Here we introduce a different approach: using probabilistic physics-informed machine learning to model the spatiotemporal patterns of total sea-level change—barystatic, steric, and glacial isostatic adjustment—in a unified framework. We call our novel, uncertainty-aware methodology SELENE: SEa-Level Emulator by Neural Estimation. We test SELENE by comparing the emulated total sea-level change during the 21st century (2000-2025) with the satellite altimetry data, demonstrating the capability of our novel approach to explain the observations’ variance by up to 80% (R-squared metric, across all grid cells) on multidecade forecast horizon. In addition, benchmarking against the existing models reveals a 10-fold reduction in computational complexity. Hence, SELENE presents a new class of models for fast, efficient, and accurate sea-level change modeling and projection, with application in geodesy and many other areas of Earth sciences.</jats:p>