Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Graph Neural Network as Computationally Efficient Emulator of Ice-sheet and Sea-level System Model (ISSM)

Younghyun Koo, Maryam Rahnemoonfar

TL;DR

This work addresses the computational bottleneck of CPU-only ISSM by introducing a GPU-based Graph Convolutional Network (GCN) emulator trained on 20-year Pine Island Glacier simulations. The GCN ingests inputs $x$, $y$, time, and basal melting rate to predict outputs $v_x$, $v_y$, and ice thickness $h$, using five graph convolutional layers with 128 features and distance-weighted neighbor aggregation. Fidelity analyses show correlations exceeding $0.996$ (and up to near $0.998$ reported) with ISSM for velocity and thickness, with GCN outperforming a Fully Convolutional Network baseline. In terms of performance, the GPU-based GCN achieves about $33.8\times$ speedup over CPU ISSM, enabling rapid exploration of basal-melting-rate scenarios and future sea-level impacts, thereby offering a scalable tool for climate-change impact studies on Antarctic and Greenland ice sheets.

Abstract

The Ice-sheet and Sea-level System Model (ISSM) provides solutions for Stokes equations relevant to ice sheet dynamics by employing finite element and fine mesh adaption. However, since its finite element method is compatible only with Central Processing Units (CPU), the ISSM has limits on further economizing computational time. Thus, by taking advantage of Graphics Processing Units (GPUs), we design a graph convolutional network (GCN) as a fast emulator for ISSM. The GCN is trained and tested using the 20-year transient ISSM simulations in the Pine Island Glacier (PIG). The GCN reproduces ice thickness and velocity with a correlation coefficient greater than 0.998, outperforming the traditional convolutional neural network (CNN). Additionally, GCN shows 34 times faster computational speed than the CPU-based ISSM modeling. The GPU-based GCN emulator allows us to predict how the PIG will change in the future under different melting rate scenarios with high fidelity and much faster computational time.

Graph Neural Network as Computationally Efficient Emulator of Ice-sheet and Sea-level System Model (ISSM)

TL;DR

This work addresses the computational bottleneck of CPU-only ISSM by introducing a GPU-based Graph Convolutional Network (GCN) emulator trained on 20-year Pine Island Glacier simulations. The GCN ingests inputs , , time, and basal melting rate to predict outputs , , and ice thickness , using five graph convolutional layers with 128 features and distance-weighted neighbor aggregation. Fidelity analyses show correlations exceeding (and up to near reported) with ISSM for velocity and thickness, with GCN outperforming a Fully Convolutional Network baseline. In terms of performance, the GPU-based GCN achieves about speedup over CPU ISSM, enabling rapid exploration of basal-melting-rate scenarios and future sea-level impacts, thereby offering a scalable tool for climate-change impact studies on Antarctic and Greenland ice sheets.

Abstract

The Ice-sheet and Sea-level System Model (ISSM) provides solutions for Stokes equations relevant to ice sheet dynamics by employing finite element and fine mesh adaption. However, since its finite element method is compatible only with Central Processing Units (CPU), the ISSM has limits on further economizing computational time. Thus, by taking advantage of Graphics Processing Units (GPUs), we design a graph convolutional network (GCN) as a fast emulator for ISSM. The GCN is trained and tested using the 20-year transient ISSM simulations in the Pine Island Glacier (PIG). The GCN reproduces ice thickness and velocity with a correlation coefficient greater than 0.998, outperforming the traditional convolutional neural network (CNN). Additionally, GCN shows 34 times faster computational speed than the CPU-based ISSM modeling. The GPU-based GCN emulator allows us to predict how the PIG will change in the future under different melting rate scenarios with high fidelity and much faster computational time.
Paper Structure (11 sections, 1 equation, 5 figures, 2 tables)

This paper contains 11 sections, 1 equation, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. ISSM simulation for the Pine Island Glacier
  4. Graph Convolutional Network (GCN)
  5. Fully Convolutional Network (FCN)
  6. Results
  7. Experiment set-up
  8. Fidelity of the graph neural network emulators
  9. Computational performance of the graph neural network emulators
  10. Sensitivity to basal melting rate
  11. Conclusion

Figures (5)

  • Figure 1: (a) Location of the Pine Island Glacier (PIG) in Antarctica; (b) Ice velocity, (c) surface elevation, and (d) ice thickness of the PIG acquired from NASA MEaSUREs data.
  • Figure 2: Schematic illustration of the graph convolutional network (GCN) emulator.
  • Figure 3: Maps of the 20-year-averaged ice velocity for different basal melting rates
  • Figure 4: Maps of the 20-year-averaged ice thickness for different basal melting rates
  • Figure 5: Changes in mean ice velocity and ice thickness of the PIG under different melting rates (0, 20, 40, and 60 m/year).