Prognostic Validation of a Neural Network Unified Physics Parameterization

• Published in: Geophysical research letters Vol. 45; no. 12; pp. 6289 - 6298
• Main Authors: Brenowitz, N. D., Bretherton, C. S.
• Format: Journal Article
• Language: English
• Published: Washington John Wiley & Sons, Inc 28.06.2018; American Geophysical Union (AGU)
• Subjects: Artificial intelligence; Atmosphere; Atmospheric models; Climate; Climate models; cloud‐resolving model; Computer applications; Computer science; Computer simulation; Computers; cumulus parameterization; Digital imaging; Learning algorithms; Machine learning; Model testing; Modelling; neural network; Neural networks; Parameterization; Physics; Rain clouds; Resolution; Robustness (mathematics); Scale (ratio); Scientists; Simulation; single‐column model; Weather
• ISSN: 0094-8276; 1944-8007
• DOI: 10.1029/2018GL078510

Weather and climate models approximate diabatic and sub‐grid‐scale processes in terms of grid‐scale variables using parameterizations. Current parameterizations are designed by humans based on physical understanding, observations, and process modeling. As a result, they are numerically efficient and interpretable, but potentially oversimplified. However, the advent of global high‐resolution simulations and observations enables a more robust approach based on machine learning. In this letter, a neural network‐based parameterization is trained using a near‐global aqua‐planet simulation with a 4‐km resolution (NG‐Aqua). The neural network predicts the apparent sources of heat and moisture averaged onto (160 km)2 grid boxes. A numerically stable scheme is obtained by minimizing the prediction error over multiple time steps rather than single one. In prognostic single‐column model tests, this scheme matches both the fluctuations and equilibrium of NG‐Aqua simulation better than the Community Atmosphere Model does. Plain Language Summary Computer‐based atmospheric models predict future conditions based on physical laws. Unfortunately, current computers are not powerful enough to simulate these physical laws exactly and must represent the atmosphere as a 3‐D digital image, with a pixel size, also known as resolution, of 100–200 km. Scientists typically approximate any processes that occur below this size, such as rain clouds, using idealized physical models. Unfortunately, the ability of these models to represent the atmosphere's true complexity is limited by a scientist's imagination. The recent availability of large data sets, computational resources, and developments in computer science allows a different approach based on machine learning. In this article, we train one popular kind of machine learning model, known as neural network, to approximate the small‐scale processes in climate model based on a high‐resolution atmospheric simulation. To produce an accurate prediction, we had to reduce the prediction error of many days, rather than just a few hours. The resulting neural network matches the high‐resolution simulation better than a state of the art traditional climate model does, a heartening result which suggests that machine learning can be blended with traditional approaches to produce better predictions of future weather and climate. Key Points A neural network‐based unified physics parameterization is trained on a near‐global aqua‐planet simulation from a cloud‐resolving model A numerically stable scheme is trained by minimizing the prediction error accumulated over multiple time steps Prognostic single‐column simulations with the neural network scheme closely match the target data