Spatio‐Temporal Super‐Resolution Data Assimilation (SRDA) Utilizing Deep Neural Networks With Domain Generalization.

Deep learning has recently gained attention in the atmospheric and oceanic sciences for its potential to improve the accuracy of numerical simulations or to reduce computational costs. Super‐resolution is one such technique for high‐resolution inference from low‐resolution data. This paper proposes a new scheme, called four‐dimensional super‐resolution data assimilation (4D‐SRDA). This framework calculates the time evolution of a system from low‐resolution simulations using a physics‐based model, while a trained neural network simultaneously performs data assimilation and spatio‐temporal super‐resolution. The use of low‐resolution simulations without ensemble members reduces the computational cost of obtaining inferences at high spatio‐temporal resolution. In 4D‐SRDA, physics‐based simulations and neural‐network inferences are performed alternately, possibly causing a domain shift, that is, a statistical difference between the training and test data, especially in offline training. Domain shifts can reduce the accuracy of inference. To mitigate this risk, we developed super‐resolution mixup (SR‐mixup)–a data augmentation method for domain generalization. SR‐mixup creates a linear combination of randomly sampled inputs, resulting in synthetic data with a different distribution from the original data. The proposed methods were validated using an idealized barotropic ocean jet with supervised learning. The results suggest that the combination of 4D‐SRDA and SR‐mixup is effective for robust inference cycles. This study highlights the potential of super‐resolution and domain‐generalization techniques, in the field of data assimilation, especially for the integration of physics‐based and data‐driven models. Plain Language Summary: One challenge in the Earth sciences is the simulation of atmospheric and oceanic processes with high accuracy or at high spatio‐temporal resolution. The neural network, a model developed in the field of artificial intelligence, has recently gained attention for its potential to learn the relationship between any variables from a sufficient amount of data. For example, a neural network can be trained to make high‐resolution images from low‐resolution images. This process is called super‐resolution. We propose the incorporation of a neural network for super‐resolution into an atmospheric or ocean model that performs numerical simulations at low resolution. Since the atmospheric or ocean model is of low resolution, the proposed framework is computationally efficient while still allowing for high‐resolution results with the aid of the neural network. Furthermore, to make the inference accurate, observation data are fused with atmospheric or ocean predictions in the neural network. In real‐world applications, it is important to make predictions robust to noise from various factors. Accordingly, we developed a new technique where certain noise is added during the training of the neural network. The approach proposed here can efficiently compute high‐resolution predictions of the atmosphere and ocean without compromising robustness. Key Points: We propose a neural network‐based scheme for data assimilation and spatio‐temporal super‐resolution using physics‐based simulationsA data‐augmentation technique was developed to improve the robustness of inference in the fusion of data‐driven and physics‐based modelsA test with idealized ocean jets showed that the proposed approach efficiently infers high‐resolution results with ensuring the robustness [ABSTRACT FROM AUTHOR]