Porting OptClim Optimisation system to ARCHER2
ARCHER2-eCSE04-07
PI: Prof Simon Tett (University of Edinburgh)
Co-Is: Dr Andrew Schurer (University of Edinburgh), Dr Coralia Cartis (University of Oxford), Dr Dan(i) Jones (British Antarctic Survey), Dr Maria Val Martin (University of Sheffield)
Technical staff: Dr Mike Mineter (University of Edinburgh), Dr Magnus Hagdorn (University of Edinburgh)
Subject Area:
Some weather processes can’t be fully resolved in maths, and have to be parameterised. Cloud physics is prominent in this.
Climate and Earth system models do not have perfect information, nor perfect representation of physical processes as they cannot represent all physical scales from µm to Mm. Consequently they cannot be expected to be produce exact representations of reality. When used to reproduce observations (rather than for in silico experiments to gain understanding) discrepancy can come in many forms. This includes model error due to missing or incorrectly modelled processes. Uncertainty in small scale processes leads to large scale system uncertainty. Also any single model simulation will not have the same realisation in time of natural variability (such as simulating El Nino events).
We have made available two software systems to enhance the ways in which researchers can respond to uncertainty in models. Firstly, in models some processes are parametrised rather than modelled with detailed maths and data, for example those concerning small-scale processes in clouds, that have large-scale consequences for the radiation balance. We have ported OptClim2 software and extended it to permit researchers to optimise parameters in the earth system models CESM, MITgcm-based models and UKESM. Secondly we have developed software for “particle filtering” (PF) in which an ensemble of UKESM runs is used to track historical data. The systems are currently being trialled by the developers, with UKESM.
Documentation, code and status are held on GitHub. The development was complicated by the need to communicate from ARCHER2 to PUMATEST (the NCAS-provided server that runs UKESM workflow). This was achieved with a previous package we had developed, GeosMeta, a service providing a RESTful interface to a Mongo Database.
It is a pleasure to acknowledge timely support from both ARCHER2 and from NCAS.
Information about the code
OptClim2:
Code for OptClim2 is in the ARCHER2 github branch with documentation in its wiki.
Each use for research requires added code to be tailored (to calculate simulated observables, choosing parameters…). An initial exemplar is in OptClim_UKESM github
Particle Filtering for UKESM:
Use of this requires also added code tailored for the specific goals of the research
The OptClim and the Particle Filtering code will continue to be maintained from the University of Edinburgh by this research group and be available on GitHub for others to use. We expect that others will make changes (and commit those back) to GitHub.
Once the new PUMA service, from which UKESM models will be run, is available it will be closer in terms of security to ARCHER2. It is expected that some relatively minor work will be needed to simplify the codes to benefit from this change.