A Framework for Deep Learning Emulation of Numerical Models With a Case Study in Satellite Remote Sensing.

Area of Science:

• Earth System Science
• Computational Science
• Machine Learning

Background:

• Physics-based numerical models are crucial for Earth system insights but face computational limitations.
• High-resolution simulations are desired but often computationally prohibitive.
• Surrogate models are used to approximate complex numerical models.

Purpose of the Study:

• To investigate the efficacy of deep learning (DL) emulation for approximating Earth system model simulations.
• To assess DL emulation's computational efficiency and accuracy compared to traditional surrogate models.
• To explore DL's potential for capturing complex spatiotemporal dependencies in Earth system processes.

Main Methods:

• Developed and applied a deep learning emulation framework.
• Utilized a case study in satellite-based remote sensing for validation.
• Compared the performance of DL emulation against a traditional surrogate model.

Main Results:

• Deep learning emulation successfully reproduced simulation results with acceptable accuracy.
• DL emulation demonstrated comparable or superior computational efficiency.
• The DL approach effectively captured complex processes and spatiotemporal dependencies.

Conclusions:

• Deep learning emulation is a credible and efficient method for approximating complex Earth system models.
• DL offers a viable solution to overcome computational constraints in Earth science research.
• Advancements in DL performance support its broader application for high-resolution Earth system simulations.