Title: Scalable Gaussian process for computer model emulation and uncertainty quantification

Speaker: Mengyang Gu, UCSB

Abstract: Computer experiments are ubiquitously used in science and engineering, yet the computational complexity can prohibit its applications for many large-scale systems. In this talk, we will introduce the Gaussian process emulator as a surrogate model for approximating computer model simulations and uncertainty quantification. The new results concern emulating computer models with massive coordinates, high-dimensional inputs and functionals. To overcome the computational bottleneck in the Gaussian process, we will review the stochastic differential equation representation of Gaussian processes with a Matern covariance, and applications of Kalman filter, as an  example of marginalization that results in an exact, computationally efficient alternative. Our extensions include modeling incomplete matrix observations of correlated data, estimating interaction kernels and forecasting trajectories in particle dynamics.  We have developed software packages in R and MATLAB that implement the Gaussian process emulator and some of the computationally scalable alternatives. Applications include emulating the TITAN2D model of pyroclastic flows, ground deformation simulation by COMSOL Multiphysics, molecular dynamics simulations, and cell alignment processes.



Bio: Mengyang Gu is an assistant professor at Department of Statistics and Applied Probability at UC Santa Barbara. He obtained his PhD in Statistical Science at Duke University. He is enthusiastic in developing computational tools and software packages to integrate physical models and statistical learning approaches for accelerating simulations with controlled errors, automating model evolution, and ensuring computational scalability, efficiency, robustness and convergence of the algorithms. He received the SIAM activity group on uncertainty quantification early career prize in 2022, for his contribution to the analysis and estimation of Gaussian process emulators.