Unraveling Tropical Weather Secrets: CNNs Illuminate Madden-Julian Oscillation Predictability

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jan 15 2024

In an article published in the journal Nature, researchers investigated the predictability of the Madden-Julian Oscillation (MJO), a key tropical weather pattern, using a 1200-year simulation from the community Earth system model version 2 (CESM2).

Employing a convolutional neural network (CNN) trained on the simulation and fine-tuned with observations, the study enhanced MJO prediction skill to around 25 days. Explainable artificial intelligence (XAI) methods identified precipitable water anomalies in the Indo-Pacific warm pool as crucial precursors for accurate MJO prediction, emphasizing the importance of realistic moisture dynamics.

Background

The MJO is a crucial tropical weather phenomenon with a significant impact on high-impact weather events globally. Despite advancements in dynamical models, operational models face challenges in accurately predicting MJO, hindering forecasts of weather events beyond two weeks. This study addressed the limited understanding of MJO predictability sources, a key barrier to improving prediction skills.

While AI and machine learning (ML) techniques have been applied to MJO prediction, their skills remained below operational forecasts, and the sources of MJO predictability for extended lead times (>15 days) were inadequately explored. Researchers employed a 1200-year climate simulation from the CESM2 to develop a robust regression CNN model for MJO prediction. The CNN was fine-tuned using transfer learning with observations to mitigate CESM2 biases.

The model demonstrated skillful forecasting of observed MJO, revealing atmospheric water vapor anomalies as the primary source of predictability within a 3-week lead time. This approach filled gaps by leveraging a long-term simulation dataset, applying transfer learning to address model biases, and quantifying the impact of input field changes on predicted output values, contributing to a more comprehensive understanding of MJO predictability.

Methods

In this study, a 1200-year-long pre-industrial simulation with the CESM2 was utilized to investigate the predictability of the MJO. Various atmospheric variables, including outgoing longwave radiation (OLR), zonal winds (U250 and U850), precipitable water (PW), and surface temperature (ST), were considered. The dataset was pre-processed to create anomalies and normalized for input to a CNN. The CNN was initially trained using CESM2 simulation data for different forecast lead times (one to 30 days).

Transfer learning was then applied to fine-tune the CNN with observational data, enhancing its ability to predict MJO by addressing model biases. The CNN demonstrated improved prediction skills, surpassing a score of 0.6 at a 15-day forecast lead time with more than 400 years of training data. Transfer learning with observational data further refined the model.

The authors introduced the signed-contribution (SC) map, an XAI method, to quantify the influence of each input variable on the CNN's output. The SC map, incorporating sign information, provided insights into the direction and magnitude of the relationship between input variables and MJO prediction. Additionally, a bivariate-mean-square-difference (BMSD) analysis assessed the impact of individual input variables on the CNN's predictions, revealing the contribution of each variable to the overall predictability of MJO.

These methodologies contributed to a comprehensive understanding of MJO predictability, utilizing a long-term simulation dataset, incorporating transfer learning to address model biases, and employing XAI methods to interpret the CNN's predictions.

Results

The researchers evaluated the performance of a CNN-based MJO prediction model, trained on a 1200-year CESM2 simulation and fine-tuned with observational data using transfer learning. The CNN achieved a bivariate correlation (BCOR) skill score above 0.5 up to approximately 25 days of forecast lead time, outperforming previous ML models and comparable to several dynamical models, except for the European Centre for Medium-Range Weather Forecasts (ECMWF) model. Transfer learning significantly improved the model's skill, particularly for weeks two and three, addressing systematic biases in CESM2.

Investigating the MJO's predictability source, the authors employed an XAI method, the SC map, revealing that PW anomalies played a dominant role in predictability for up to three weeks. Removing PW significantly decreased the model's skill, emphasizing the importance of moisture dynamics. Other variables, such as zonal winds and ST, also contributed, with their impacts varying with forecast lead time. Lag-correlation maps showed that the model captured moisture anomalies' predictive features, demonstrating the ability to use moisture memory for predictions.

The BMSD analysis quantified the relative importance of input variables, indicating PW's substantial influence, followed by upper-level zonal winds and ST. The researchers highlighted the significance of PW as the primary source of MJO predictability, providing insights into the specific regions influencing predictions. The results suggested the potential for further enhancing early forecast lead times (< 10 days) and underscored the need for a realistic representation of moisture dynamics in MJO prediction models.

Discussion

The researchers underscored the potential of ML models in weather and climate prediction, showcasing the CNN-based MJO prediction model's remarkable skill of approximately 25 days, outperforming previous ML models. While still behind some dynamical models, the results suggested further improvements can be achieved by leveraging high-quality, high-volume data from general circulation models (GCMs) that better simulate the MJO.

The findings emphasized the significance of large-scale moisture anomalies as the primary source of MJO predictability for up to three weeks, supporting the theoretical 'moisture mode' framework. It also suggested the applicability of ML models for revealing the nature of target phenomena and highlighted the potential for using different deep learning techniques and multi-model simulations to enhance prediction skills and understand associated uncertainties. The discussion underscored ML models' efficiency in computation time compared to traditional numerical approaches, opening avenues for combined use with dynamical models in weather and climate prediction.

Conclusion

In conclusion, this research demonstrated the efficacy of a CNN-based MJO prediction model trained on a 1200-year CESM2 simulation, fine-tuned with observations through transfer learning. The model outperformed previous ML models and aligned with dynamical models. Moisture anomalies, particularly PW, emerged as the dominant source of MJO predictability for up to three weeks. Leveraging longer simulation data and exploring diverse ML techniques offer avenues for further improvements. The researchers enhanced the understanding of MJO predictability sources, highlighting the potential of ML models for comprehensive weather and climate predictions.