Uncertainty quantification and propagation through complex chains of computational models
- Research Team
- Stephen Gow
- Investigators
- Dave Woods
This project will explore how predictions can be made and assessed through complex chains of computer models. For example, consider predicting casualties from the release of a biological or chemical agent. Modelling such outcomes requires linking predictions of meteorology, atmospheric dispersion, sensor properties and dose response. Each model will be subject to uncertainties including uncertain inputs, uncertain tuning parameters, and uncertain physical mechanisms.
Understanding the reliability and accuracy of the overall predictions of casualties requires understanding how the uncertainties from the individual models will propagate and combine. This project will develop the necessary methodology for construction of accurate statistical emulators, or surrogates, of chains of models to reduce computational cost; data fusion from a variety of models and data sources of different fidelities; and the necessary algorithms to allow computationally feasible Bayesian inference for multi-model chains.
The research will be motivated by, and demonstrated on, multi-model chains from Dstl which are used for hazard response, hazard management and government procurement programmes.
Categories
Algorithms and computational methods: statistical analysis
Programming languages and libraries: C++, R
Transdisciplinary tags: NGCM