Workshop: Artificial Intelligence for Weather and Climate Modelling
Abstract
This workshop, the first of its kind dedicated entirely to the theme of applications of AI for weather and climate modelling, is motivated by the need to improve the accuracy of regional simulations in global weather and climate models. Increasingly society is becoming dependent on such models for a variety of reasons. Not least, they are now used for predictions on monthly, seasonal and decadal timescales, helping provide warnings of drought, flood and periods of severe weather. On multi-decadal timescales, such models are being used to provide guidance on infrastructure investment to adapt to changing extremes of weather and climate, and to evaluate the regional consequences of geoengineering proposals. However, due principally to the inherent inaccuracy of the subgrid parametrisation process, weather and climate models still suffer from systematic biases when compared with observations. These biases are often as large as the signals which the models attempt to predict or simulate. A route to improving the accuracy of models is to increase resolution so that key physical processes (such as deep convection, orographic gravity wave drag and mixing by mesoscale eddies in the ocean) can be represented with the proper laws of physics, rather than with approximate parametrisation formulae.However, this is computationally challenging and will require substantial improvements in computational efficiency, even with the type of exascale hardware envisaged in the coming years.AI can help us achieve the required computational efficiency. By training neural nets on output from sub-grid parametrisations, it may be possible to replace these parametrisations in production/operational runs of a weather/climate model without degrading skill but at considerable computational speed up. Research (performed under an ONR grant) has already indicated that much of the model can be run using half-precision numerics without degradation. This suggests that AI-based parametrisations running on GPUs with halfprecisioncapability may be an exceptionally efficient way to represent current expensivesub-grid parametrisations (and modules of Earth System complexity more generally). By reinvesting the computational savings thus made into the dynamical cores of weather/climate models, and utilising next-generation exascale computing, it may finally be possible to reach the goal of an accurate global cloud-resolved model in the coming 5 years or so. A spin-off of this work is also improved efficiency 4D variational data assimilation systems as linearised neural nets provide an efficient way to represent complex parametrisation processes in the optimisation procedures in data assimilation.Going beyond this, there is an additional potential benefit of AI for weather and climate modelling ~ as a tool for developing new parametrisations - by training the AI software either on observations or on very high resolution limited-area model output. However, this second application is more controversial. Not least how do we ensure any such AI-based parametrisations respond properly in conditions which it has not met in the training dataset ~ for example when the climate model is run into the future with atmospheric carbon dioxide far higher than they are now? Or alternatively when the model is run over past paleoclimatic periods far removed from today~s conditions?
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Aug 15, 2019
- Source ID
- N000141912548
Entities
People
- Tim Palmer
Organizations
- Office of Naval Research
- United States Navy
- University of Oxford