Arctic weather and climate: from mechanisms to forecasts

Since the mid-20th Century, the Arctic is warming about two times faster than the rest of the world. This is widely known as Arctic Amplification (AA) and has drawn lots of attention recently. Apart from a dominant impact on the local weather and climate, its influence has been extended beyond the polar region through significant modification on the circulation in both the atmosphere and ocean. In order to identify the drivers of Arctic climate change and the connections between the Arctic and global climate system, it is important to understand the Arctic weather and climate.

The Arctic weather and climate system is manifested by the cryosphere. Given the essential role of sea ice and the key factors that regulate the variability of sea ice, it is of vital importance to shed new light on the energy budget, as well as the atmosphere-ocean interaction in the perspective of the energy budget in the Northern Hemisphere. A better quantification of the Arctic sea ice variability and the Arctic energy budget will give us a better understanding of the mechanisms behind Arctic climate warming. This in turn would help improve the weather forecasts for the Northern Hemisphere.

This thesis, I mainly focus on the variations and forecasting of the Arctic sea ice. Given the close relation between the Arctic energy budget and sea ice variability, an emphasis is placed on the meridional energy transport (MET). In chapter 2, I quantified meridional energy transport in the atmosphere (AMET) and ocean (OMET) at subpolar latitudes using the latest methods and six reanalysis data sets. An intercomparison of the results from the chosen reanalysis products indicates that although the mean transport in all data sets agrees well, the spatial distributions and temporal variations of AMET and OMET differ substantially among the reanalysis data sets. Our study confirms that the analyzed reanalysis products are useful for the diagnostics of energy transport. However, beyond interannual timescales, the results must be interpreted with caution, especially when studying variability and interactions between the Arctic and midlatitudes.

The analysis of energy transport is extended to the whole energy budget and the interactions between atmosphere and ocean in the Northern Hemisphere in chapter 3. Based on our findings in chapter 2, I examined the compensation between heat transport variations in the atmosphere and ocean. Different from studies in the past, which were mostly based on numerical climate models, we provided new insights into the so-called Bjerknes compensation using multiple reanalysis products. Our study shows that Bjerknes compensation is present at mid-latitudes in the Northern Hemisphere from interannual to decadal time scales. We found that the response of the mean flow to OMET variability leads to the Bjerknes compensation. It is the shift of the Ferrel cell at midlatitudes at decadal time scales in winter, which is driven by the eddy momentum flux that causes the compensation. Our findings are different from some experiments with numerical climate models, which attribute the compensation to the variation of transient eddy transports in response to the changes of OMET at multidecadal time scales.

On the basis of physical insights into the energy transport and sea ice variability shown in chapter 2 and 3, the study continues in the direction of sea ice forecasts and a series of experiments of extended range sea ice forecasts with novel deep neural networks are presented in chapter 4. Inspired by the rapid developments in deep learning techniques, in this chapter I proposed Convolutional Long Short Term Memory Networks (ConvLSTM) to forecast sea ice in the Barents Sea at weather to sub-seasonal time scales. The architecture of this neural network is designed to exploit the covariances between different variables, including spatial and temporal relations, which is suitable for the prediction of spatial-temporal sequential data. Using reanalysis products, we demonstrated that ConvLSTM is able to learn the variability of the Arctic sea ice and can forecast regional sea ice concentration skillfully at weekly to monthly time scales. In general, forecasts from ConvLSTM outperform those with climatology, persistence, and a statistical model, and they are comparable to the forecasts from operational sub-seasonal to seasonal weather forecast systems. Furthermore, we demonstrated that the ConvLSTMs are able to preserve the physical consistency between predictors and predictands in their forecasts. A sensitivity test aiming at an evaluation of the impact of different predictors on the quality of forecasts shows that the surface energy budget components have a significant impact on the predictability of sea ice at weather time scales. Our findings indicate that such data-driven methods with deep neural networks are promising tools for enhancing operational Arctic sea ice forecasting.

Our exploration is further extended to probabilistic deep learning in chapter 5. Motivated by the need for uncertainty quantification in weather forecasts and the limit on uncertainty estimation reflected by previous studies using deterministic neural networks, we explored probabilistic deep neural networks for weather forecasting. By replacing fixed weights with distributions following the methodology in Bayesian deep learning (BDL), we turned normal Long-Short Term Memory neural networks (LSTM) into Bayesian Long-Short Term Memory neural networks (BayesLSTMs). With an aim to understand the characteristics of BDL within a simplified dynamical system that represents the essence of midlatitude atmospheric dynamics, we used BayesLSTMs to forecast output from the Lorenz 84 system with seasonal forcing. We showed that forecasts with the BayesLSTM can stay close to the attractor of the Lorenz model and concluded that they represent the nonlinear relations between each components in this simplified atmospheric circulation system. We further demonstrated that the BayesLSTMs are able to produce reliable probabilistic forecasts and address uncertainties relevant to weather forecasting. Our study indicates that BDL is an easy and fast solution for probabilistic weather forecast and is promising to enhance weather forecasting capabilities at short to medium-range timescales.

Finally, a synthesis of this thesis and an outlook for suggested future research are shown in chapter 6.