Bayesian Treed Multivariate Gaussian Process With Adaptive Design: Application to a Carbon Capture Unit

Abstract

Computer experiments are widely used in scientific research to study and predict the behavior of complex systems, which often have responses consisting of a set of nonstationary outputs. The computational cost of simulations at high resolution often is expensive and impractical for parametric studies at different input values. In this article, we develop a Bayesian treed multivariate Gaussian process (BTMGP) as an extension of the Bayesian treed Gaussian process (BTGP) to model the cross-covariance function and the nonstationarity of the multivariate output. We facilitate the computational complexity of the Markov chain Monte Carlo sampler by choosing appropriately the covariance function and prior distributions. Based on the BTMGP, we develop a sequential design of experiment for the input space and construct an emulator. We demonstrate the use of the proposed method in test cases and compare it with alternative approaches. We also apply the sequential sampling technique and BTMGP to model the multiphase flow in a full scale regenerator of a carbon capture unit.