The Next Generation of Surrogate Modelling in Environmental Science
9th - 11th July 2018
The Welcome Centre, Lancaster University
The workshop brought together statisticians, computer scientists and environmental modellers to discuss how we could work more effectively together, with a particular focus on surrogate models.
Models that are used to simulate different parts of the earth system are essential tools for overcoming environmental challenges (e.g. climate change or air pollution). However, these models are often slow to run, meaning that they are limited in terms of:
(i) the level of spatial and temporal resolution they can operate at
(ii) the number of runs that can be carried out for comprehensive quantification of model uncertainty
For example, a typical global atmospheric chemistry-climate model may take 24 hours of computer time to complete a one-year simulation, half of which is used in solving the extensive system of differential equations needed to represent the chemistry. A solution is to replace the slowest part of the model with a simpler statistical model called a Gaussian process emulator. However, carrying this out presents a major challenge: the outputs of the part of the complex model that we are emulating are typically multivariate. The covariances between the different output variables cannot be ignored – as is commonly done in current uses of emulators – because these outputs are used for further calculations in other parts of the model.
The workshop welcomed almost 40 researchers from academia and industry, half of whom were external to the university. It involved 12 excellent talks which were followed by two half days of some very useful discussions. We agreed on two activities following on from this workshop:
(i) the planning and implementation of a ‘surrogate modelling competition’ with prizes to be donated by an industry partner
(ii) a review paper to discuss how we can bring together applied science modellers (who want to use surrogate models) and those who currently use surrogate models (e.g. statisticians)
If you would like to be involved in any of these follow-on activities or you would like more details about the workshop, please e-mail Ed at email@example.com.