Comparing neural operator based surrogate models on glacier dynamics prediction.

Assigned Session: #AGM28: Generic Meeting Session

Abstract ID: 28.7385 | Accepted as Poster | Poster | 2025-02-28 12:45 - 14:15 | Ágnes‐Heller‐Haus/Small Lecture Room

Mamta K C (0)
Köstler, Harald (2), Fürst, Johannes J. (1)
Mamta K C ((0) Friedrich-Alexander-Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Bavaria, DE)
Köstler, Harald (2), Fürst, Johannes J. (1)

(0) Friedrich-Alexander-Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Bavaria, DE
(1) Institute of Geography, Friedrich-Alexander Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Germany
(2) Department of Computer Science, Friedrich-Alexander-Universität Erlangen-Nürnberg, Cauerstraße 11, 91058, Erlangen, Germany

(1) Institute of Geography, Friedrich-Alexander Universität Erlangen-Nürnberg, Wetterkreuz 15, 91058, Erlangen, Germany
(2) Department of Computer Science, Friedrich-Alexander-Universität Erlangen-Nürnberg, Cauerstraße 11, 91058, Erlangen, Germany

Categories: Cryospheric Processes
Keywords: Surrogate Models, Glacier Dynamics, Neural Operators

Categories: Cryospheric Processes
Keywords: Surrogate Models, Glacier Dynamics, Neural Operators

Large-scale and long-term simulations of ice-dynamic glacier evolution are computationally expensive using traditional numerical solver strategies. Surrogate models, trained on data simulated from traditional solvers or directly integrating the governing Full-Stokes equations offer a computationally efficient alternative. Convolutional Neural networks(CNNs) have been successfully applied to accelerate prediction while maintaining adequate accuracy. However, the CNN architecture exhibits limited generalization abilities as it operates on finite dimensional Euclidean spaces, thus working for particular discretization or resolution. Unlike CNNs, neural operators can map between functions in infinite-dimensional spaces making them resolution more invariant and better at generalization. This study therefore explores the use of neural operator-based surrogate models to predict glacier velocity and evaluates the performance of different architectures with respect to their generalization abilities.


NAME:
Small Lecture Room
BUILDING:
Ágnes‐Heller‐Haus
FLOOR:
0
TYPE:
Lecture Hall
CAPACITY:
200
ACCESS:
Only Participants
ADDITIONAL:
TBA
FIND ME:
>> Google Maps