ML approaches to accelerate the spin-up phase of climate models: A software engineering perspective.
Host
Alan Edelman
## Title:
ML approaches to accelerate the spin-up phase of climate models: A software engineering perspective.
## Abstract:
The ICCS collaborates with two VESRI sites (M2Lines and CALIPSO) toward accelerating the spin-up phase of climate models using data driven methods. The term 'spin-up' is the process by which traditional climate models reach equilibrium; a necessary step to run a numerical experiment forwards in time. This process, be it a land surface model or ocean model, and a set of initial conditions, can take up to 98% of the running time of a typical model experiment, which may consume months of computing resources on a supercomputer. This is a major bottleneck in integrating ever increasing observational data (a big model - big data dilemma).
Recently, climate researchers have identified various ways to use data-driven machine learning based acceleration tools (MLA) to reduce the computational demand of traditional spin-up methods. In this talk, we will give an overview of the SPINAcc and Spinup-NEMO projects from a software engineering perspective, focusing on the project redesign and refactoring that ICCS engineers are currently undertaking, as well efforts towards optimising the spin-up acceleration tool which maximises the computational time savings and minimises biases in the ML predictions.
===========
## Bio:
Matthew Archer is a senior research software engineer in the ICCS based at the University of Cambridge. He currently specialises in machine learning for climate science applications and is generally interested in the accelerated spin-up of climate models using data driven techniques. He previously contributed to the development of the finite element software library FEniCS, after completing a PhD in Computational Physics from the same university specialising in Solid Mechanics.
ML approaches to accelerate the spin-up phase of climate models: A software engineering perspective.
## Abstract:
The ICCS collaborates with two VESRI sites (M2Lines and CALIPSO) toward accelerating the spin-up phase of climate models using data driven methods. The term 'spin-up' is the process by which traditional climate models reach equilibrium; a necessary step to run a numerical experiment forwards in time. This process, be it a land surface model or ocean model, and a set of initial conditions, can take up to 98% of the running time of a typical model experiment, which may consume months of computing resources on a supercomputer. This is a major bottleneck in integrating ever increasing observational data (a big model - big data dilemma).
Recently, climate researchers have identified various ways to use data-driven machine learning based acceleration tools (MLA) to reduce the computational demand of traditional spin-up methods. In this talk, we will give an overview of the SPINAcc and Spinup-NEMO projects from a software engineering perspective, focusing on the project redesign and refactoring that ICCS engineers are currently undertaking, as well efforts towards optimising the spin-up acceleration tool which maximises the computational time savings and minimises biases in the ML predictions.
===========
## Bio:
Matthew Archer is a senior research software engineer in the ICCS based at the University of Cambridge. He currently specialises in machine learning for climate science applications and is generally interested in the accelerated spin-up of climate models using data driven techniques. He previously contributed to the development of the finite element software library FEniCS, after completing a PhD in Computational Physics from the same university specialising in Solid Mechanics.