Tuning complex computer codes to data and optimal designs

Modern scientific researchers often use complex computer simulation codes for theoretical investigations. We model the response of computer simulation code as the realization of a stochastic process. This approach, design and analysis of computer experiments (DACE), provides a statistical basis for analysing computer data, for designing experiments for efficient prediction and for comparing computer-encoded theory to experiments. An objective of research in a large class of dynamic systems is to determine any unknown coefficients in a theory. The coefficients can be determined by "tuning" the computer model to the real data so that the tuned code gives a good match to the real experimental data. Three design strategies for computer experiments are considered: data-adaptive sequential A-optimal design, maximum entropy design and optimal Latin-hypercube design. The following "code tuning" methodologies are proposed: nonlinear least squares, joint MLE, "separated" joint MLE and Bayesian method. The performance of these methods have been studied in several toy models. In the application to nuclear fusion devices, a cheaper emulator of the simulation code (BALDUR) has been constructed, and the transport coefficients were estimated from data of two tokamaks (ASDEX and PDX). Tuning complex computer codes to data using some statistical estimation methods and a cheap emulator of the code along with careful designs of computer experiments, with applications to nuclear fusion devices, is the topic of this thesis.