Machine dependence as a source of uncertainty in climate models: The HadGEM3-GC3.1 CMIP Preindustrial simulation

When the same weather or climate simulation is run on different High Performance Computing (HPC) platforms, model outputs may not be identical for a given initial condition. While the role of HPC platforms in delivering better climate projections is often discussed in literature, attention is mainly...

Full description

Bibliographic Details
Main Authors: Guarino, Maria-Vittoria, Sime, Louise C., Schroeder, David, Lister, Grenville M. S., Hatcher, Rosalyn
Format: Text
Language:English
Published: 2019
Subjects:
Archer
Sea ice
Online Access:https://doi.org/10.5194/gmd-2019-83
https://www.geosci-model-dev-discuss.net/gmd-2019-83/
Description
Summary:When the same weather or climate simulation is run on different High Performance Computing (HPC) platforms, model outputs may not be identical for a given initial condition. While the role of HPC platforms in delivering better climate projections is often discussed in literature, attention is mainly focused on scalability and performance rather than on the impact of machine-dependent processes on the numerical solution. At the same time, machine dependence is an overlooked source of uncertainty when it comes to discussing the model spread observed within the Coupled Model Intercomparison Projects (CMIP). Here we investigate the impact of machine dependence on model results and quantify, for a selected case study, the magnitude of the uncertainty. We consider the Preindustrial (PI) simulation prepared by the UK Met Office for the forthcoming CMIP6. We compare key climate variables between PI control simulations run on the UK Met Office supercomputer and the ARCHER HPC platform. Discrepancies strongly depend on the timescale. Decadal means show substantial differences of up to 0.2 °C for global mean air temperature, 1 W/m 2 for TOA outgoing longwave flux and 1.2 million km 2 for Southern Hemisphere sea ice area. However, on multi-centennial timescales the differences are not significant and the long-term statistics of the two runs are similar. Differences between the two simulations can be linked to variations in the strongest modes of climate variability. In the Southern Hemisphere, this results in large SST anomalies where ENSO teleconnection patterns are expected that can reach 0.6 °C (and SNR > 1) even on centennial timescales.

Similar Items