Evaluating statistical consistency for the ocean component of earth system models using physics informed convolutional autoencoder

Model verification is a crucial step in the development and optimization of Earth system models (ESMs). Recently, a deep learning approach in evaluating statistical consistency for the atmosphere component of ESMs (A-ESM-DCT) has been proven effective in model verification. However, due to the longer timescales required for the ocean to adjust compared to the atmosphere, achieving an efficient method for evaluating the consistency of ocean models necessitates extending model integration time while reducing ensemble simulation size. Traditional deep learning models, such as those in A-ESM-DCT, often lack robustness and fail to guarantee convergence in small data regimes. We introduce a deep learning consistency test for the ocean component of ESMs, referred to as O-ESM-DCT. The O-ESM-DCT is based on a physics-informed convolutional autoencoder (PIConvAE) model to achieve rapid convergence with significantly fewer simulations and adopts the data-fidelity and physical-driven losses as an indicator to determine the consistency. The O-ESM-DCT is applicable for evaluating the consistency of Community Earth System Model in heterogeneous computing environments as well as modifications to the compiler, numbers of computing cores, and model parameters. Our O-ESM-DCT offers an effective and objective solution for ensuring the reliability of the development and optimization processes for the ocean component of ESMs on high-performance computing systems.