High-fidelity physics simulations, such as the Energy Exascale Earth System Model (E3SM), are generating ever increasing quantities of data. In the near future, it will be infeasible to store the full computation after the simulation. To facilitate the training of complex machine learning models in the future, there is a glaring need for in-situ inference algorithms. In this work, we focus on the need for spatio-temporal inference models which are (i) highly scalable, (ii) distributed, (iii) able to capture small-scale, high-resolution structures, and (iv) longrange global structure. One possible approach is to leverage algorithms for fast approximate Gaussian process regression, such as the sparse variational Gaussian process (SVGP). In this work, we present a Hierarchical SVGP (HSVGP) model in which local models capture small-scale structures in-situ, while large-scale structures are captured and communicated by a global SVGP. We also consider the simpler approach in which independent, parallel sparse Gaussian processes are fit to each local region with no additional global information. Using both synthetic data and E3SM model output, we show the success of the HSVGP approach in some settings, but also the surprisingly good performance of the independent approach.
|