Learning Parameters from Data: Calibration, Inverse Problems, and Model SIAM UQ mini symposium, Friday 9 April, EPFL, Lausanne.
Alex Diaz, Marco Iglesias and myself have organised a mini symposium at the SIAM UQ conference, which will be in Lausanne this year.
** Session Abstract**
Complex numerical models have become so ubiquitous in science, that nowadays the conclusions drawn from their use influence critical and expensive decisions in science, engineering and public policy. The predictive power of such models relies on how well they are calibrated to experimental data. The main aim of this minisymposium is to bring together experts in calibration and inverse problem modelling. Relevant topics for include: robust calibration techniques, inverse problem methodologies, Bayesian model updating, history matching, data assimilation, model discrepancy, amongst others. Theoretical and computational developments are of interest, with particular focus on numerical efficiency. Organizer: Alejand
Faced with situations that we wish to improve, we need to formulate some kind of model of how our controllable inputs to the situation affect the outputs. Whether we are selecting the best inputs to result in the kids going to bed on time, or making a quiet aircraft engine, modelling the situation is important. Models may be analytical, physics-based or empirical, or a mixture. While analytical and physics-based models can be fast to run, they may not be able to predict true-life outcomes. Physical experiments are often expensive and display all the negatives/excitement of true-life - primarily uncertainty. Statistical modelling methods are useful for combining data of varying lineage into a useful model that can be used to predict outcomes, and so optimise situations. In this talk I will show the development of such a model. It is a musculoskeletal/hydrodynamic model of the underwater fly-kick of an Olympic swimmer, based on experimental, physics-based and analytical data. Statistical model based optimization techniques are then used to develop swimming techniques for optimal trade-offs between power and thrust. The techniques are applicable in all areas of science where inputs can be modified to affect outputs which are to be improved.
The speakers are
- Alejandro Diaz (University of Liverpool)- History matching with structural reliability methods
- Ian Vernon (Durham University) - Iterative History Matching for Computationally Intensive Inverse Problems
- James Hensman (University of Lancaster) - Efficient Inverse problems with Gaussian process surrogates via entropy search
- Zaid Sawlan (KAUST) - Bayesian inference for linear parabolic PDEs with noisy boundary conditions