Simulation-based inference of surface accumulation and basal melt rates of an Antarctic ice shelf from isochronal layers

Guy Moss, Vjeran Višnjević, Olaf Eisen, Falk M. Oraschewski, Cornelius Schröder, Jakob H. Macke, Reinhard Drews

Abstract

The ice shelves buttressing the Antarctic ice sheet determine the rate of ice-discharge into the surrounding oceans. Their geometry and buttressing strength are influenced by the local surface accumulation and basal melt rates, governed by atmospheric and oceanic conditions. Contemporary methods quantify one of these rates, but typically not both. Moreover, information about these rates is only available for recent time periods, reaching at most a few decades back since measurements are available. We present a new method to simultaneously infer the surface accumulation and basal melt rates averaged over decadal and centennial timescales. We infer the spatial dependence of these rates along flow line transects using internal stratigraphy observed by radars, using a kinematic forward model of internal stratigraphy. We solve the inverse problem using simulation-based inference (SBI). SBI performs Bayesian inference by training neural networks on simulations of the forward model to approximate the posterior distribution, therefore also quantifying uncertainties over the inferred parameters. We validate our method on a synthetic example, and apply it to Ekström Ice Shelf, Antarctica, for which independent validation data are available. We obtain posterior distributions of surface accumulation and basal melt averaging over up to 200 years before 2022.

Introduction

Ice shelves act as vital "corks" that hold back the Antarctic ice sheet from sliding into the ocean. To predict future sea-level rise, scientists need to know two things: how much snow falls on top (accumulation) and how much the ocean melts them from below (basal melt).

The challenge is that current satellite and field measurements only go back a few decades. Furthermore, most existing methods can only calculate one of these rates at a time, often by assuming the other is already known. This paper tackles the "inverse problem": given the current internal structure of the ice (layers seen via radar), can we work backward to find the historical accumulation and melt rates that created those layers? Because the physics involved is complex and the "likelihood" of any specific scenario is mathematically difficult to calculate, traditional statistical methods often fail or are too slow.

Methodology

The authors use a cutting-edge approach called Simulation-Based Inference (SBI), specifically a variant known as Neural Posterior Estimation (NPE).

Schematic of the ice shelf model showing the Mass Budget method versus the Internal Reflection Horizon (IRH) method used in this study.

The methodology consists of four primary components:

  1. The Forward Model: A fast kinematic model simulates how ice layers (isochrones) form over time based on flow speed, snow accumulation, and basal melt.
  2. Simulation-Based Inference: Instead of solving complex equations directly, the team ran the forward model 190,000 times with varying random inputs to create a comprehensive "what-if" library.
  3. Neural Networks: They trained a neural network (specifically Normalizing Flows) to learn the relationship between simulated radar layers and input rates. A 1D Convolutional Neural Network serves as an "embedding net" to compress complex radar data into a manageable format.
  4. Amortization: Once trained, the network can instantly process real radar data to provide a "posterior distribution"—a range of likely values for accumulation and melt, along with the probability of each.

Data and Validation

The framework was tested using both synthetic and real-world datasets:

  • Synthetic Data: A 2D "flow tube" simulation created with icepack software validated the model's ability to recover known "ground truth" values.
  • Ekström Ice Shelf (EIS): Real-world radar data from a 130 km transect in Dronning Maud Land, Antarctica, was used for application.
  • Validation: Results were compared against independent snow accumulation measurements from the Kottas Traverse (1996–2023) and existing satellite/airborne basal melt maps.

Key Insights and Results

The SBI approach offers several advantages over traditional glaciological methods:

  • Simultaneous Inference: This is one of the first methods to solve for both accumulation and basal melt at the same time without needing one as a fixed input.
  • Historical Reconstruction: Deeper layers represent averages over longer periods. By analyzing layers at different depths, the authors estimated mass balance averaging over roughly 40 to 200 years.
  • Age Estimation: The model can estimate the age of internal ice layers without needing to drill expensive ice cores. For example, it estimated one layer to be ~84 years old, matching historical expectations.
  • Uncertainty Quantification: Unlike "best-guess" models, this approach provides a rigorous measure of uncertainty, which is crucial for climate modeling.

On synthetic tests, the model recovered true rates within a 95% confidence interval. On the Ekström Ice Shelf, inferred snow accumulation matched independent ground measurements remarkably well. The model successfully captured high melt rates (up to 4 meters per year) near the grounding line and showed how melting stabilizes further out to sea.

Limitations and Future Work

While promising, the current model assumes a "steady-state," meaning the ice shelf's shape and velocity haven't changed drastically over the last 200 years. If the ice shelf is currently thinning or accelerating rapidly, these results might be biased. Furthermore, the study focuses on 2D flow lines; real-world ice flow is 3D and more complex. Future work aims to incorporate "transient" signals to move beyond the steady-state assumption and scale the SBI approach to higher resolutions.

Summary updated

Abstract

Abstract The ice shelves buttressing the Antarctic ice sheet determine the rate of ice-discharge into the surrounding oceans. Their geometry and buttressing strength are influenced by the local surface accumulation and basal melt rates, governed by atmospheric and oceanic conditions. Contemporary methods quantify one of these rates, but typically not both. Moreover, information about these rates is only available for recent time periods, reaching at most a few decades back since measurements are available. We present a new method to simultaneously infer the surface accumulation and basal melt rates averaged over decadal and centennial timescales. We infer the spatial dependence of these rates along flow line transects using internal stratigraphy observed by radars, using a kinematic forward model of internal stratigraphy. We solve the inverse problem using simulation-based inference (SBI). SBI performs Bayesian inference by training neural networks on simulations of the forward model to approximate the posterior distribution, therefore also quantifying uncertainties over the inferred parameters. We validate our method on a synthetic example, and apply it to Ekström Ice Shelf, Antarctica, for which independent validation data are available. We obtain posterior distributions of surface accumulation and basal melt averaging over up to 200 years before 2022.
BibTeX
@article{moss2025simulation,
  author = {Guy Moss and Vjeran Višnjević and Olaf Eisen and Falk M. Oraschewski and Cornelius Schröder and Jakob H. Macke and Reinhard Drews},
  title = {{Simulation-based inference of surface accumulation and basal melt rates of an Antarctic ice shelf from isochronal layers}},
  journal = {Journal of Glaciology},
  year = {2025},
  doi = {10.1017/jog.2025.13},
}
Figure