Skip to Content

ESMs Latent Space Exploration for Uncertainty Quantification and Spatiotemporal Downscaling

Active Dates 9/1/2022-6/30/2025
Program Area Atmospheric System Research
Project Description
Forecasting the rapidly changing climate is one of the main goals of scientific investigators and legislators to assess, quantify, and make informed decisions about climate risks. Drought, megafires, rising sea levels, eroding shorelines, loss of wetland, and deterioration of water quality have all led to widespread concerns about climate change in the world. Advanced Earth System Models (ESMs) have been extensively used for such analyses through developing quantitative understanding of the multiscale dynamics of the climate in response to various stimuli and/or perturbations. Despite significant advancements, ESMs are still computationally very expansive and cannot accurately forecast future variations in global and/or regional climate systems due to two grand challenges: (i) the presence of large epistemic (i.e., knowledge) and aleatoric (i.e., data) uncertainties in scenario testing and predictions; and (ii) the lack of resolution to resolve mesoscale processes and mechanistic integration of the physics that occur on temporal scales from seconds to centuries or spatial scales from micrometer to macrometer. 

The overarching objective of this proposal is to relax the requirements of many Earth system models for uncertainty quantification and to establish an intelligent data framework for reliable interpolation across a wide range of spatial and temporal scales. To this end, we propose to leverage advanced unsupervised machine learning techniques comprising of dimensionality reduction and clustering algorithms to select representative ESMs (RESMs) from a larger set without a priori knowledge while preserving the uncertainty domain. This leads to improving and accelerating uncertainty quantification and risk analysis through the extraction of principal criteria. We will then decompose the RESMs to extract latent interdependencies across the scales in the forms of hybrid dynamical-statistical downscaling models that map important fine-scale variations in climate models without incurring additional computing expenses. 

We strongly believe that our work fulfills the Department of Energy’s call to minimize processing and computational requirements while improving the prediction capability of rapidly changing climates. This “Developing Capabilities” proposal would remove significant research barriers at Texas State University (TXST) and would better position TXST to participate in future Biological and Environmental Research (BER) program’s solicitations in the areas of Earth system modeling and data management.
Award Recipient(s)
  • Texas State University San Marcos (PI: Faroughi, SalahAldin)