Emulation and calibration of computer models with unknown constraints

Computer simulations have been widely used to mimic the behavior of complex physical systems. For some simulators, valid outcomes can only be obtained when the input satisfies some unknown constraints. In this thesis, new methodologies are developed around such computer models for emulation, experimental design and model calibration. This work is motivated by binary population synthesis simulation codes that study characteristics of binary black hole (BBH) mergers. Though it is computationally inexpensive to perform a single simulation with these simulators, the amount of computation needed to fully explore the input space is prohibitive. Fast and parallelizable surrogate model that combines a local GP classifier and a local GP is introduced in a Bayesian framework to emulate the simulator. Based on this emulator, a sequential design procedure is then developed to help plan computer experiments in ways that take into account both the unknown constraints and outcome uncertainty. Next, by combining the proposed emulator and field observations of binary black holes, inference on model parameters, which is part of the simulator input, is made under different scenarios with exact inference and approximate Bayesian computation. Lastly, statistical power of designed experiments with multi-level factors are investigated.