In-Situ Spatial Inference on Climate Simulations with Sparse Gaussian Processes

As extreme-scale physics simulation becomes increasingly memory and storage expensive, the ability to access full simulation data for statistical analysis is becoming increasingly limited. The capability to perform in-situ statistical inference of state variables is becoming increasingly important for the comprehensive utilization of the huge amounts of information generated by these simulations. In this work, we report the first results fitting scalable Gaussian process regression to the state information of an expensive simulation in-situ. For this, spatial regression of upper atmosphere temperature data was performed using Julia coupled to the E3SM climate model. The resulting sparse Gaussian process model shows strong predictive performance using a small number of representative observations. These results provide the backbone for more general in-situ spatial inference with Gaussian process models in complex physics simulations.