Hybrid physics-ML & parameterization★ Highlight

Vertically Recurrent Neural Networks for Sub‐Grid Parameterization

Peter Ukkonen, Matthew Chantry

Journal of Advances in Modeling Earth Systems · 01 June 2025 · ECMWF, University of Oxford

Why it matters. Shows that recurrent networks improve ML parameterization emulators, with online IFS radiation tests and an honest speed comparison. Relevant to hybrid modelling.

In one sentence. Shows recurrent networks emulate IFS radiation, gravity waves and moist physics with stable autoregressive runs, matching ResNets offline and running more efficiently.

Abstract

Abstract Machine learning has the potential to improve the physical realism and/or computational efficiency of parameterizations. A typical approach has been to feed concatenated vertical profiles to a dense neural network. However, feed‐forward networks lack the connections to propagate information sequentially through the vertical column. Here we examine if predictions can be improved by instead traversing the column with recurrent neural networks (RNNs) such as Long Short‐Term Memory (LSTMs). This method encodes physical priors (locality) and uses parameters more efficiently. Firstly, we test RNN‐based radiation emulators in the Integrated Forecasting System. We achieve near‐perfect offline accuracy, and the forecast skill of a suite of global weather simulations using the emulator are for the most part statistically indistinguishable from reference runs. But can radiation emulators provide both high accuracy and a speed‐up? We find optimized, state‐of‐the‐art radiation code on CPU generally faster than RNN‐based emulators on GPU, although the latter can be more energy efficient. To test the method more broadly, and explore recent challenges in parameterization, we also adapt it to data sets from other studies. RNNs outperform reference feed‐forward networks in emulating gravity waves, and when combined with horizontal convolutions, for non‐local unified parameterization. In emulation of moist physics with memory, the RNNs have similar offline accuracy as ResNets, the previous state‐of‐the‐art. However, the RNNs are more efficient, and more stable in autoregressive semi‐prognostic tests. Multi‐step autoregressive training improves performance in these tests and enables a latent representation of convective memory. Recently proposed linearly recurrent models achieve similar performance to LSTMs.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.