Role of the ocean for fast atmospheric evolution revealed by machine learning

TL;DR

This work explores how model forecast errors relate to properties of the air-sea interface, and infer what ocean information these atmospheric emulators are able to derive from atmospheric data alone, and what they cannot, to inform how future machine learning models should use ocean information on short timescales.

Abstract

There have recently been many efforts to create machine learnt atmospheric emulators designed to replace physical models. So far these have mainly focused on medium-range weather forecasting, where these `Machine Learnt Weather Prediction' (MLWP) models can outperform leading operational forecasting centres. However, because of this focus on shorter timescales, many of these emulators ignore the effects of the ocean, and take no ocean variables as inputs. We hypothesise that such MLWP models have learnt a best-guess of the evolution of the atmosphere, by implicitly inferring ocean conditions from atmospheric states, with no access to ocean data. Turning this limitation into a strength, we use it as a means to study the role of the oceans on the evolution of the atmosphere. By exploring how model forecast errors relate to properties of the air-sea interface, we infer what ocean information these atmospheric emulators are able to derive from atmospheric data alone, and what they cannot. This highlights the regions and processes through which the ocean independently influences the atmosphere on fast timescales. We perform this analysis for GraphCast, finding clear relationships between air-sea properties and the forecast errors over the ocean, including clear seasonal effects. We then explore what this reveals about GraphCast's internal representation of the ocean. In addition to understanding real-world ocean-atmosphere interactions, this analysis provides guidance for improving forecast skill and physical realism in MLWP models, and for informing how future machine learning models should use ocean information on short timescales.

Figures (5)

• Figure 1: Spearman correlation between GraphCast's T2m error at 24hrs lead time and variables at the sea surface (a) SST for June, (b) SST-T2m for June, (c) SST for December, (d) SST-T2m for December. Hatching indicates where the correlation is not significant according to a two-tailed t-test at the 95% confidence level. Note that the colorbar increases exponentially.
• Figure 2: Inferred areas of ocean-driven dynamics, using the technique of lagged correlations between the observed SST anomaly and mean surface heat flux anomaly (a) over June months and (b) over December months. Correlation is shown only for the points that have significant correlation at zero lag, and where the correlation does not change sign between a lead/lag of 72 hours. Red (blue) colouration indicates where there is significant positive (negative) correlation that does not change sign with lagging. Note the colorbar increases exponentially.
• Figure 3: Top row: Spearman correlation between GraphCast's errors and (a) SST climatology for June (b) SST climatology minus T2m climatology for June. Bottom row: Spearman correlation between GraphCast's errors aggregated to a monthly level and (c) monthly SST and (d) monthly SST-T2m, for June.
• Figure 4: Linear model coefficients (using a lasso model with $\alpha=0.0001$) for June for (a) the SST input field and (b) the SST-T2m input field. In this case the model is fitted with all variables simultaneously, so where the model coefficients are zero indicates a point where that variable is not important for the prediction.
• Figure 5: Difference in AWRMSE by grid cell between the baseline model with $\alpha=100$ and the optimal model with $\alpha=0.001$ (a) AWRMSE difference and (b) percentage AWRMSE difference for each grid cell. Negative (blue) values indicate where the error of the optimal model is lower (i.e. where the optimal linear model has good predictive performance).