Abstract
Estimating surface accumulation and basal melt rates of Antarctic ice shelves is critical for predicting global sea-level rise. However, these rates are difficult to measure simultaneously over long timescales. This paper introduces a novel framework using Simulation-Based Inference (SBI) to "unmix" these variables by treating internal ice layers—visible via radar—as a historical record. By training a neural posterior estimator on simulations from a kinematic ice-flow model, the authors reconstruct nearly 200 years of mass balance history for the Ekström Ice Shelf with rigorous uncertainty quantification. The method demonstrates that the shelf has remained in a relatively stable state for two centuries and provides a scalable approach for processing large-scale radar datasets.
Introduction
Antarctic ice shelves serve as critical "plugs" that regulate the flow of inland ice into the ocean. Their stability is governed by a delicate mass balance: snow accumulation on the surface and basal melting driven by ocean heat. Understanding how these rates fluctuate over decadal and centennial timescales is essential for climate modeling.
Historically, glaciologists have struggled to solve for both accumulation and melt rates simultaneously, often assuming one is constant to solve for the other. While ice cores offer high-resolution historical data, they are expensive, point-based measurements. This research seeks to turn internal reflection horizons (isochrones)—widely available through radar surveys—into a continuous, spatially distributed archive of past climate and ocean conditions.
Methods
The authors frame the reconstruction of ice shelf history as an inverse problem: determining the accumulation and melt rates that produced the observed geometry of internal ice layers.
Forward Modeling and SBI
The framework utilizes a kinematic "tracer method" based on the Shallow Shelf Approximation (SSA) to simulate how ice flows and deforms over time. To overcome the computational intractability of traditional likelihood-based methods, the team employed Simulation-Based Inference (SBI).
- Neural Posterior Estimation (NPE): Approximately 190,000 simulations were generated to build a training set. A Normalizing Flow (Neural Spline Flows) was trained to learn the posterior distribution of parameters given the data.
- Embedding Network: A 1D Convolutional Neural Network (CNN) was used to compress high-dimensional radar data into informative summary statistics.
- Noise Modeling: A power-spectral density noise model was implemented to bridge the "sim-to-real" gap, accounting for measurement errors and sub-kilometer variability.
Datasets
The model was validated using synthetic data and applied to newly acquired ground-based radar measurements (2021–2023) from the Ekström Ice Shelf (EIS). Validation was performed against independent snow accumulation measurements from the Kottas traverse stake arrays, with geometry data sourced from BedMachine Antarctica and ITS_LIVE.
Key Results
- High-Fidelity Reconstruction: The posterior predictive distributions reconstructed observed radar layers with significantly higher accuracy than the prior, reducing Root Mean Square Error (RMSE) from 11.8m to 4.6m for a 30m deep layer.
- Validation Success: Inferred surface accumulation rates closely matched independent physical measurements, successfully capturing a distinct accumulation peak roughly 20 km from the grounding line.
- Historical Reach and Age Estimation: The model successfully dated radar layers back to approximately 188 years ago. By comparing results across different depths, the authors found consistent spatial trends, providing strong evidence that the Ekström Ice Shelf has remained in a relatively stable steady state for the last two centuries.
- Amortized Inference: Once trained, the neural network performs inference on new radar data almost instantaneously, allowing for the rapid processing of massive datasets without re-running expensive simulations.
Discussion and Future Work
While the current framework assumes a steady state, it provides a robust foundation for future glaciological studies. The authors acknowledge that for rapidly changing ice shelves, such as those in West Antarctica, the model must be adapted to handle transient, time-varying flow. Future improvements may include moving toward a "Lagrangian" parameterization to better couple surface accumulation to local topography and exploring spectral representations to handle higher-resolution spatial data as the number of parameters increases.
Summary updated