Every year, from June to September, the South Asian monsoon season delivers copious rainfall to more than one billion individuals residing in the Indian subcontinent. 

The rainfall exhibits periodic fluctuations, with certain weeks experiencing precipitation ranging from 1 to 4 inches, while other weeks are predominantly arid. Anticipating the timing of these arid and rainy seasons is crucial for agricultural and municipal planning, facilitating farmers' awareness of the optimal time to gather crops and aiding city officials in preparing for potential flooding. Although weather forecasts are often reliable for short-term periods of one or two days, accurately predicting weather conditions for a week or a month in advance is highly challenging.

The researchers created a new machine-learning-based forecast that has been shown to more accurately predict South Asian monsoon rainfall 10 to 30 days in advance. It significantly improves over current forecasts using numerical modelling rather than artificial intelligence. Understanding monsoon behaviour is crucial because it is a significant atmospheric characteristic in the global climate.

Weather forecasting

Weather forecasting is challenging due to various instabilities in the atmosphere. These instabilities include continuously heating the atmosphere from the Earth's surface, resulting in colder and denser air above hotter and less dense air. Additionally, the uneven heating of the atmosphere and the rotation of the Earth contribute to its instability. These instabilities result in a tumultuous state where the inaccuracies and uncertainties in estimating the atmosphere's behaviour rapidly amplify, rendering it exceedingly challenging to forecast events in the distant future.

Research information

The paper is titled "Improved subseasonal prediction of South Asian monsoon rainfall using data-driven forecasts of oscillatory modes." Eviatar Bach, the Foster and Coco Stanback Postdoctoral Scholar Research Associate in Environmental Science and Engineering, led the research. Bach works in the laboratories of Tapio Schneider, the Theodore Y. Wu Professor of Environmental Science and Engineering and JPL senior research scientist, as well as Andrew Stuart, the Bren Professor of Computing and Mathematical Sciences. 

In addition to Bach, coauthors are:

• V. Krishnamurthy, Center for Ocean-Land-Atmosphere Studies, George Mason University, Fairfax, VA 22030
• Safa Mote, Fariborz Maseeh Department of Mathematics and Statistics, Portland State University, Portland, OR 97201. Department of Atmospheric and Oceanic Sciences and Institute for Physical Science and Technology, University of Maryland, College Park, MD 20742
• Jagadish Shukla, Department of Atmospheric, Oceanic and Earth Sciences, George Mason University, Fairfax, VA 22030
• A. Surjalal Sharma, Department of Astronomy, University of Maryland, College Park, MD 20742. (Deceased January 5, 2024).
• Eugenia Kalnay, Department of Atmospheric and Oceanic Sciences and Institute for Physical Science and Technology, University of Maryland, College Park, MD 20742
• Michael Ghil, Geosciences Department and Laboratoire de Météorologie Dynamique (CNRS and Institut Pierre-Simon Laplace), École Normale Supérieure and Paris Sciences et Lettres University, Paris, France 75005. Department of Atmospheric and Oceanic Sciences, University of California, Los Angeles, CA 90095. Department of Mathematics, Imperial College London, London SW7 2BX, United Kingdom

Existing models

Presently, cutting-edge models employ numerical modelling, which consists of computer simulations of the atmosphere generated by applying physical equations that describe fluid motion. Due to disorder, the most common time for predicting large-scale weather is around ten days. Forecasting the long-term average atmosphere or climate behaviour is also feasible. However, numerical models have encountered difficulties when foretelling the weather within two weeks to several months.

Monsoon intraseasonal oscillations (MISOs)

The South Asian monsoons are characterized by heavy rainfall in cycles of violent downpours followed by periods of dry weather. The term used to describe these cycles is monsoon intraseasonal oscillations (MISOs). The researchers incorporated a machine-learning element into the existing cutting-edge numerical models. It enabled them to collect data on the MISOs and improve their ability to forecast rainfall on the illusive two-to-four-week timescale. The resultant model significantly enhanced the correlations between the predictions and observations by up to 70%.

Conclusion

The researchers assert that this approach might also be utilized for other intraseasonal oscillations, such as the Madden–Julian oscillation, where data-driven predictions have proven highly accurate. In a broader sense, this study showcases the effectiveness of integrating dynamic and data-based models for forecasting Earth system behaviour.

Recent research on hybrid forecasting combining different methods has shown promising results. The field of solely data-driven full-field forecasts on subseasonal-to-seasonal timescales is also emerging. The researchers anticipate merging dynamical and data-driven forecasts for real-time prediction and data assimilation as machine learning improves its ability to forecast weather and climate. 

Source: https://www.pnas.org/doi/10.1073/pnas.2312573121

Image source: Unsplash