Pulong Ma

Research

My research is generally focused on developing Bayesian statistical and machine learning methods and theory for modeling spatial data, including geostatistical data and physical simulations. My current work falls into three directions.

1. Bayesian tree-based methods for high-dimensional spatial data

Spatially correlated data exhibit complicated structures such as heterogeneity and local regime changes. Accounting for these structures in modeling has been a fundamental challenge in spatial statistics. Modern measurement technologies have resulted in increasingly large data collected over domains with more than two dimensions, which makes the problem harder still and calls for new statistical methods that address the trade-off between model flexibility and computational scalability. Currently, I work on Bayesian tree methods and multiscale modeling approaches for Gaussian processes and spatial point processes that are suited to large-scale data-analytic problems arising in high-dimensional space. (Supported by NSF DMS-2348163)

2. Hierarchical mixture for space-time processes

Space-time processes often exhibit asymmetric structures between space and time. For instance, the covariance between today’s temperature in Ames and yesterday’s temperature in Cincinnati differs from that between yesterday’s temperature in Ames and today’s temperature in Cincinnati. Geophysical processes often evolve jointly in space and time with interactions, resulting in complex structures including long-range dependence and highly nonstationary and asymmetric covariance structures. I develop a hierarchical mixture framework that produces valid space-time covariance models with these features, and study their theoretical properties as well as their practical use in prediction. (Supported by NSF DMS-2348154)

3. Bayesian UQ and simulation-based inference for complex physical systems

Complex physical systems are studied through computational models (or simulators) built from physical knowledge, which are imperfect representations of the real-world process. To estimate physical parameters and predict the real-world process, I develop computationally scalable statistical methods for high-dimensional simulator outputs and simulation-based inference approaches for settings where the likelihood is intractable or inference is too expensive, but simulator runs are available. This work is driven by collaborations with domain experts in remote sensing, atmospheric science, ocean science, and engineering. (Supported by U.S. DOE DE-SC0026595)




Last modified by Pulong Ma on Oct 16, 2026