Glacier mass balance modeling using a Long Short-Term Memory network

Glacier mass balance (MB) is a key indicator of climate change and a central driver of glacier evolution, yet most glaciers worldwide lack long-term in-situ measurements. For estimating glacier MB, data-driven models provide a complementary alternative to traditional numerical approaches by learning empirical relationships between climate forcing, topography, and MB directly from observations. Here, we develop a recurrent neural network based on a Long Short-Term Memory architecture within the Mass Balance Machine (MBM) framework and evaluate its ability to predict winter and annual point surface MB across the Swiss Alps. MBM is trained on 30,000 observations from 30 glaciers and tested on eight glaciers excluded from training to assess spatial generalization. MBM predicts winter and annual MB with high accuracy on unseen glaciers (RMSE values of 0.35 and 0.78 m w.e., respectively). Its recurrent structure enables learning temporal dependencies, thereby improving the representation of seasons with strong accumulation or ablation. Beyond point predictions, MBM generates spatially distributed MB maps that closely resemble reference products derived from observations, capture MB gradients, and produce glacier-wide mass changes consistent with geodetic estimates. Monthly outputs further show that MBM reproduces the seasonal transition from winter accumulation to summer ablation with realistic timing and magnitude. These results show that a recurrent neural network can recover key characteristics of glacier MB dynamics from observations and that the learned relationships transfer effectively across glaciers with different climatic and topographic settings. The demonstrated generalization skill highlights the potential of MBM for application in regions with limited direct measurements.