Schmitt et al. (2026) AGILE v0.1: The Open Global Glacier Data Assimilation Framework

Identification

• Journal: Geoscientific model development
• Year: 2026
• Date: 2026-02-12
• Authors: Patrick Schmitt, Fabien Maussion, Daniel Goldberg, Philipp Gregor
• DOI: 10.5194/gmd-19-1301-2026

Research Groups

• Department of Atmospheric and Cryospheric Sciences, University of Innsbruck, Innsbruck, Austria
• School of Geographical Sciences, University of Bristol, Bristol, United Kingdom
• School of Geosciences, University of Edinburgh, Edinburgh, United Kingdom
• Meteorologisches Institut, Ludwig-Maximilians-Universität München, München, Germany

Short Summary

This paper introduces AGILE v0.1, an open global glacier data assimilation framework that uses a time-dependent variational method with automatic differentiation to efficiently optimize glacier bed topography and distributed ice volume, demonstrating significant improvements over initial guesses in idealized experiments.

Objective

• To develop and evaluate AGILE v0.1, an open global glacier data assimilation framework, for transient calibration of glacier models.
• To recover glacier bed topography and distributed ice volume in 2020 through transient calibration, based on dynamical simulations from 1980, using idealized experiments.
• To demonstrate AGILE's ability to improve upon existing bed inversion methods and reconstruct dynamically consistent glacier states.

Study Configuration

• Spatial Scale: Global glacier modeling, tested on four synthetic glaciers (Aletsch, Artesonraju, Baltoro, Peyto) representing different climates and sizes, using a 1.5D flowline representation.
• Temporal Scale: Transient simulations from 1980 to 2020 (40 years), with target observations in 2000 and a mass balance period from 2000 to 2020.

Methodology and Data

• Models used: AGILE v0.1 (Open Global Glacier Data Assimilation Framework), a time-dependent variational method inspired by 4D-Var, built on a PyTorch reimplementation of the OGGM (Open Global Glacier Model) flowline glacier evolution model. Optimization uses the L-BFGS-B algorithm. First guesses were generated using OGGM's default bed inversion and the GlabTop method. A simple degree-day model was used for mass balance.
• Data sources: Synthetic glacier geometries and observations (surface elevation, total volume, geodetic mass balance) derived from OGGM simulations. Real-world data used for synthetic glacier creation included RGI v6 outlines, NASADEM, Farinotti et al. (2019) consensus ice thickness, and W5E5 climate data.

Main Results

• AGILE v0.1 efficiently optimizes glacier bed topography and distributed ice volume, requiring only a few iterations (21–27 model runs over 20 iterations) to significantly improve upon initial guesses, with minimal computational cost (approximately 1.3 seconds per iteration).
• The framework successfully recovers dynamically consistent glacier states in 2020, demonstrating substantial improvements in mean absolute differences for bed height and 2020 distributed ice volume compared to both OGGM and GlabTop first guesses.
• Reconstruction of past glacier states (e.g., 1980 distributed ice volume) without direct observations is fundamentally limited by the diffusive nature of glacier dynamics, leading to information loss over time.
• The inclusion of surface height observations (S2000) proved most influential for improving bed and volume estimates, and increasing the number of target observations generally enhanced performance.
• A broad range of regularization weights (λ from 10⁻⁴ to 10⁻¹) yielded effective results in idealized scenarios, with regularization becoming more critical for real-world, uncertain data.

Contributions

• Presents AGILE v0.1, the first application of a transient automatic differentiation (AD)-based data assimilation framework in global glacier modeling.
• Introduces an open-source framework built on a PyTorch reimplementation of the OGGM flowline model, enabling full differentiability and efficient, simultaneous optimization of multiple control variables (glacier bed elevation and distributed ice volume).
• Provides a proof-of-concept for integrating diverse, temporally distributed observations into glacier evolution models in a dynamically consistent manner, offering a flexible and computationally efficient tool for future real-world applications.
• Demonstrates the potential to improve upon existing glacier bed inversion methods and generate dynamically consistent glacier states for initialization of future projections.

Funding

• European Union’s Horizon 2020 research and innovation programme (PROVIDE, grant agreement no. 101003687)
• Austrian Climate Research Programme (ACRP) – 14th call (HyMELT-CC, grant agreement no. KR21KB0K00001)
• ESA’s “Digital Twin Component for Glaciers” project (4000146160/24/I-KE)
• NERC grants (NE/X005194/1, NE/T001607/1)
• Open Access Publication Fund of the University of Innsbruck

Citation

@article{Schmitt2026AGILE,
  author = {Schmitt, Patrick and Maussion, Fabien and Goldberg, Daniel and Gregor, Philipp},
  title = {AGILE v0.1: The Open Global Glacier Data Assimilation Framework},
  journal = {Geoscientific model development},
  year = {2026},
  doi = {10.5194/gmd-19-1301-2026},
  url = {https://doi.org/10.5194/gmd-19-1301-2026}
}