A deep-learning-based consistency test approach for Earth system models on HPC systems.

Area of Science:

• Climate Science
• Computational Science
• Artificial Intelligence

Background:

• Ensuring climate consistency is crucial for Earth System Model (ESM) development and optimization on High-Performance Computing (HPC) systems.
• Existing methods for verifying model consistency can be time-consuming and may not adequately address complexities introduced by HPC environments.

Purpose of the Study:

• To introduce an efficient and objective deep-learning-based consistency test for Earth System Models (ESMs).
• To validate the reliability of ESM simulations under various modifications within HPC environments.

Main Methods:

• Developed the Earth System Model Deep-Learning Consistency Test (ESM-DCT) utilizing an unsupervised bidirectional gate recurrent unit-autoencoder (BGRU-AE) model.
• Employed reconstruction errors from the BGRU-AE model to assess feature consistency in ESM simulation ensembles.
• Evaluated the ESM-DCT using the Community Earth System Model (CESM) on the Sunway heterogeneous system.

Main Results:

• The ESM-DCT successfully determined statistical distinguishability between new and original trusted ESM simulation ensembles.
• The test demonstrated effectiveness across heterogeneous computing environments, compilation optimization changes, and model parameter modifications.
• Reconstruction errors from the BGRU-AE model served as a reliable metric for consistency evaluation.

Conclusions:

• The ESM-DCT offers an efficient and objective approach for verifying the reliability of ESMs during development and optimization on HPC systems.
• This deep-learning tool enhances the trustworthiness of climate model simulations in complex computational settings.
• ESM-DCT facilitates robust model verification, crucial for advancing climate science research.