Geisser Lecture

Murali Haran is Professor of Statistics at Penn State University. He holds a PhD in Statistics from the University of Minnesota and a BS in Computer Science (with minors in Statistics, Mathematics, and Film Studies) from Carnegie Mellon University. His research focuses on Monte Carlo algorithms, statistical analysis of complex computer models, spatial data modeling, and interdisciplinary applications in climate science and infectious disease modeling. Haran's methodological contributions include advances in Markov chain Monte Carlo for spatial models and intractable normalizing constants, methods for high-dimensional non-Gaussian spatial data, and uncertainty quantification for simulation models. His scientific work has addressed the future of the Antarctic ice sheet, flood hazards via hydrology models, vaccination and environmental effects on infectious disease spread, and estimating disease burden from multiple imperfect data sources. He is a Fellow of the American Statistical Association and recipient of the 2015 El Shaarawi Young Researcher Award from the International Environmetrics Society. Haran served as Head of Penn State’s Department of Statistics from 2018 to 2024 and will begin a three-year term as Editor of the Journal of Agricultural, Biological, and Environmental Statistics (JABES) in 2026.

Title: A hybrid deep neural network and Gaussian process approach for computer model calibration with spatiotemporal data

Abstract: Computer models are used to simulate complex phenomena across diverse fields, including climate science, infectious disease, astronomy, and automotive crash testing. Inferring computer model parameters from observational data--a process called model calibration--is crucial because these parameters and their associated uncertainties are of direct scientific interest and heavily impact model output and predictions. Motivated by a study of water availability and drought risk, we consider calibration for a soil moisture model. This presents a challenging problem because the observations consist of high-dimensional spatiotemporal data and the model simulations can be computationally slow. Despite the extensive literature on calibration, we are u aware of existing methods that provide rigorous calibration for this class of problems. We introduce a new method that combines a deep neural network approach with a Gaussian process framework. Our method is flexible, computationally tractable, and provides uncertainty quantification; it is potentially well-suited for routine application across a wide variety of calibration problems.