Layer-Wise Relevance Propagation (LRP) offers a way to interpret the neural networks, which is generally thought as a “black box” process. We applied the LRP technique to gain a deeper understanding of the physical processes represented in subgrid convection parameterization and in predictions of Atlantic Multidecadal Variability.
Subgrid processes in global climate models are represented by parameterizations which are a major source of uncertainties in simulations of climate. In recent years, it has been suggested that machine-learning (ML) parameterizations based on high-resolution model output data could be superior to traditional parameterizations. Currently, both traditional and ML parameterizations of subgrid processes in the atmosphere are based on a single-column approach, which only use information from single atmospheric columns. However, single-column parameterizations might not be ideal since certain atmospheric phenomena, such as organized convective systems, can cross multiple grid boxes and involve slantwise circulations that are not purely vertical. Here we train neural networks (NNs) using non-local inputs spanning over 3 × 3 columns of inputs. We find that including the non-local inputs improves the offline prediction of a range of subgrid processes. The improvement is especially notable for subgrid momentum transport and for atmospheric conditions associated with mid-latitude fronts and convective instability. Using an interpretability method, we find that the NN improvements partly rely on using the horizontal wind divergence, and we further show that including the divergence or vertical velocity as a separate input substantially improves offline performance. However, non-local winds continue to be useful inputs for parameterizating subgrid momentum transport even when the vertical velocity is included as an input. Overall, our results imply that the use of non-local variables and the vertical velocity as inputs could improve the performance of ML parameterizations, and the use of these inputs should be tested in online simulations in future work.
@article{wang_non-local_2022,title={Non-{Local} {Parameterization} of {Atmospheric} {Subgrid} {Processes} {With} {Neural} {Networks}},volume={14},copyright={All rights reserved},issn={1942-2466},language={en},number={10},urldate={2024-08-17},journal={Journal of Advances in Modeling Earth Systems},author={Wang, Peidong and Yuval, Janni and O’Gorman, Paul A.},year={2022},pages={e2022MS002984},}
2023
GRL
Physical Insights From the Multidecadal Prediction of North Atlantic Sea Surface Temperature Variability Using Explainable Neural Networks
Abstract North Atlantic sea surface temperatures (NASST), particularly in the subpolar region, are among the most predictable in the world’s oceans. However, the relative importance of atmospheric and oceanic controls on their variability at multidecadal timescales remain uncertain. Neural networks (NNs) are trained to examine the relative importance of oceanic and atmospheric predictors in predicting the NASST state in the Community Earth System Model 1 (CESM1). In the presence of external forcings, oceanic predictors outperform atmospheric predictors, persistence, and random chance baselines out to 25-year leadtimes. Layer-wise relevance propagation is used to unveil the sources of predictability, and reveal that NNs consistently rely upon the Gulf Stream-North Atlantic Current region for accurate predictions. Additionally, CESM1-trained NNs successfully predict the phasing of multidecadal variability in an observational data set, suggesting consistency in physical processes driving NASST variability between CESM1 and observations.
@article{liu_physical_2023,title={Physical {Insights} {From} the {Multidecadal} {Prediction} of {North} {Atlantic} {Sea} {Surface} {Temperature} {Variability} {Using} {Explainable} {Neural} {Networks}},volume={50},issn={0094-8276},number={24},urldate={2024-08-17},journal={Geophysical Research Letters},author={Liu, Glenn and Wang, Peidong and Kwon, Young-Oh},month=dec,year={2023},pages={e2023GL106278},}
