Context-Aware Deep Learning Framework for Earth System Model Data Compression and Downscaling
Authors
Nikhil M. Pawar*, Salah A. Faroughi (salah.faroughi@txstate.edu)
Institutions
Geo-Intelligence Laboratory, Ingram School of Engineering, Texas State University–San Marcos, TX
Abstract
To conduct risk assessments and design adaptation plans against changes in climate patterns and their impacts on society, future local and regional projections of climate change hold great importance. Earth system models (ESMs) that currently run on massive supercomputers are used to generate both global and regional projections at different horizontal and vertical resolutions. These models simulate Earth’s past climate and project future scenarios, considering changes and uncertainty in atmospheric greenhouse gas emissions. ESMs are computationally expensive and require significant data resources [e.g., 1 month of simulation data using DOE’s Energy Exascale Earth System Model (E3SM) equals 14 gigabyte storage space plus memory-intensive visualization] even when operating at lower horizontal resolution (LR). For these reasons, data compression and dynamical and statistical data downscaling have been extensively practiced. In this work, the team proposes a novel two-stage framework based on deep learning to overcome this dual challenge. This framework combines an implicit neural representation (INR) and a super-resolution generative adversarial network (SRGAN). In stage 1, INR is used to compress the E3SM data by saving the weights derived from a neural network, which is trained to excessively conform to the data while maximizing overfitting. In the next stage, SRGAN is used to downscale the LR data (obtained by inferring INR) into a higher horizontal resolution (HR) to generate a more accurate representation of the ESM data for regional investigations. The primary objective of this work is to evaluate the performance of this two-stage framework in compressing and downscaling ESM data with a specific focus on surface temperature profiles derived from E3SM. Results show that the proposed framework can achieve compression gain up to four orders of magnitude with a considerable peak-to-signal ratio and generate HR profiles using the LR counterparts with significant accuracy measured based on the Learned Perceptual Image Patch Similarity metric.