Deep learning-based approaches to investigate ENSO and MJO predictability

Abstract

The emergence of deep learning (DL) has brought transformative changes to various aspects of human life and has been successfully applied in diverse domains. Leveraging its capacity to address nonlinear problems, DL is recognized as a powerful tool for learning the inherent relationships in nature, a complex system. In particular, it has been effectively employed in weather and climate forecasting, parameterization, and resolution conversion, showing improvements in performance. Nevertheless, attempts to understand the climate dynamics by interpreting DL results are limited due to the large number of parameters in DL models. Therefore, this dissertation aims to enhance our comprehension of the dynamics of climate phenomena using DL. The approach involves constructing DL models, evaluating the predictability of climate phenomena, and developing several explainable artificial intelligence (XAI) methods to elucidate the insights derived from DL results. In this regard, this thesis introduces three primary XAI methods as follows: 1) Contribution map: It quantitatively estimates the relative influence of each input variable at each grid point on the output. Its influence on the prediction is measured as the difference between the original prediction and the new prediction after zeroing the input variable, which means I remove the anomalous effects of a single grid point. 2) Signed-Contribution (SC) map: The contribution map measures only the magnitude of the influence, lacking sign information. To overcome this limitation, the signed-contribution map provides the sign of the relationship between the input and output together with the magnitude of the importance. 3) Sensitivity experiments: It estimates how much the output variable is changed to the small perturbation of the input variables. These XAI methods can identify the sole contribution or sole role of specific input variables or grid values by replacing only those components in the trained DL model and examining the differences from the original predictions. Therefore, these facilitate the investigation of the sources of predictability and its role. The target climate phenomena are the El Niño-Southern Oscillation (ENSO) and the Madden- Julian Oscillation (MJO), the most dominant modes of tropical climate variabilities, both of which influence on a global scale as well as within tropical regions. Given their substantial impacts, an accurate understanding of the dynamics of ENSO and MJO is crucial for better climate and weather prediction. Therefore, the goal of this thesis is to expand the knowledge of ENSO and MJO by assessing their predictability and sources for both phenomena based on DL and XAI methods. The ENSO, an alternation between warm (El Niño) and cold phase (La Niña), is the most prevailing tropical interannual variability, affecting global climate patterns, agriculture, and ecologies. While DL has demonstrated successful predictive capabilities for ENSO, the difficulty in interpreting DL results still prevents their applications to studies on climate dynamics. In Chapter II, I applied a convolutional neural network (CNN) to understand ENSO dynamics from long-term climate model simulations. The DL algorithm successfully predicted ENSO events with a high correlation skill (~0.82) for a 9-month lead. Based on the contribution map and sensitivity experiment, I identified three precursors of ENSO (central Pacific sea surface temperature (SST) and sea surface height (SSH), and western Pacific SSH) and investigated their physical processes with El Niño and La Niña development. In particular, it is suggested here that the roles of each precursor are asymmetric between El Niño and La Niña. In addition, the DL-based test easily reveals the actual role of the precursor, which appears to have a non-linear relationship. These results indicate that my approaches are useful for understanding physical phenomena under complex interactions. The MJO, a planetary-scale convective disturbance, is the dominant mode of tropical intraseasonal variability that interacts with many other Earth system phenomena, including high-impact weather in the midlatitude. The prediction skill of the MJO in many operational models is lower than its potential predictability, partly due to our limited understanding of its predictability source. In Chapter III, I investigate the source of the MJO predictability by combining DL. However, these MJOs have a strong seasonality, propagating eastward along the equator during the boreal winter season, but exhibiting a complex of both northward and eastward propagation during the boreal summer season. Considering this different characteristic, it is necessary to investigate the sources of MJO predictability separately by season. A CNN for MJO prediction is first trained using the 1200-year-long Community Earth System Model version 2 (CESM2) simulation and then fine-tuned using observation via transfer learning. The source of MJO predictability in the CNN is examined via the signed-contribution (SC) map. The CNN exhibits an enhanced prediction skill over previous ML models, achieving a skill level of about 24 and 26 days for the boreal winter and summer seasons, respectively. This level of performance is slightly superior or comparable to most operational models participating in the S2S project, although a few dynamical models surpass it. The SC maps highlight precipitable water (PW) anomalies over the Indo-Pacific warm pool as the primary precursors of the subsequent MJO development for 1-3 weeks forecast lead times in both seasons. For the longer lead day (> 25 days), the relative contribution of surface temperature (TS) increases, and particularly, it is enhanced during summer. These results suggest that a realistic representation of moisture dynamics is crucial for accurate MJO prediction, and also TS plays an important role in long-term memory. In addition, I especially identify the roles of PW and TS, which have a large contribution to the SC maps, in the propagation characteristics of the winter and summer MJO through XAI-based sensitivity experiments. In both seasons, a moist condition in the equatorial Indian Ocean (IO) primarily influences the eastward propagation compared to other variables. Conversely, for northward propagation during the summer, moist and warm conditions in the northern IO have comparable magnitudes of impact. Moreover, the fluctuations in moisture along the propagation direction are recognized as important factors influencing the trajectory of the MJOs. Therefore, I construct another DL-based model to predict moisture changes in the propagation direction for eastward and northward, separately. As a result, the regions with large contributions to the MJO have the same influence on moisture change. Hence, these findings suggest that variations in PW and TS in the IO are associated with moisture changes, as well as MJO-related convection. Consequently, this methodology has the advantage of being able to investigate the relative roles of specific precursors for climate phenomena. In Chapters II and III, I examined the predictability and its sources of climate variability for long-term climate simulations without external forcing, but the predictability and its sources may change with climate change due to increased carbon dioxide. Understanding how the MJO might change in response to greenhouse warming is crucial, given the potential impacts of MJO activity changes. In Chapter IⅤ, I build the DL-based MJO prediction models for current and future climates using the CESM2 large ensemble simulations and investigate the future changes in the MJO characteristics based on the XAI methods. In the future climate, while the predictability of MJO climates remains unchanged, there is a notable increase in their amplitude. The contribution map reveals that this increase in MJO amplitude is associated with an enhanced contribution of PW in the western-central Pacific (WCP), especially during future El Niño events. The enhanced contribution is influenced by the increased frequency of eastward-extended and amplified MJOs during future El Niño events. These future changes in MJO are related to the stronger response of the El Niño-related background moisture field to the same magnitude of El Niño in addition to the mean state changes under global warming. These background conditions can induce stronger horizontal moisture advection due to the strengthening of the horizontal moisture gradients. Consequently, a substantial rise in the average MJO amplitude during future El Niño plays a crucial role in the overall increase in future MJO amplitude. In summary, this thesis sheds light on how the DL can be utilized in various ways to understand climate phenomena. Diverse XAI-based approaches (i.e., contribution map, signed-contribution map, sensitivity experiments, etc.), which are first introduced in this thesis, were used to investigate the sources of ENSO and MJO predictability. Therefore, I can gain new insights that are challenging to discern using conventional linear-based methods. These methods are straightforward, intuitive, and highly scalable, making them applicable to comprehending a wide range of climate phenomena.