Machine learning in Space Weather Forecasting

Space weather refers to the branch of space physics, deals with the time varying conditions of the space surrounding the Earth. It essentially includes conditions in the magnetosphere, ionosphere, thermosphere, and exosphere. Solar wind modulates the space weather and introduces several effects on the technological systems, such as atmospheric drag in satellites, power disruption, disturbance in HF communication. Hence forecasting space weather disturbances accurately is  an active area of research.

On the other hand, lately, tremendous growth has been observed in Machine Learning (ML) and Artificial Intelligence (AI). The application of ML/AI in diverse domains yields very encouraging results in both data mining and prediction. The application of ML/AI in Space weather and solar physics study is at nascent phase, but still  provides very interesting results. Given the large volume of data captured by several satellites, it gives an excellent opportunity to explore the domain from a data science perspective. 

The main research thrust of this group is to develop and use algorithms to study and forecast the space weather effects. In different phases of the solar cycle, the automatic classification and identification of active regions, Coronal Holes and it’s corresponding solar wind are key in understanding solar physics. The prediction of solar flare, CME arrival time, solar energetic particles(SEPs) ,high-energy electron fluxes, Geomagnetic index, and Ionosphere disturbances have significant importance in both scientific and economic perspectives. We have a large and open source data set of in situ and remote observations collected over several decades through various space missions . Effective use of these data through ML/AI, may provide interesting and significant changes in space weather studies. 

Following Figure 1 shows the application of a developed algorithm for detection of coronal holes with other methods. We have already developed a Fuzzy based image processing technique to detect the coronal holes, which is one of the important sources of high solar winds. 

We are currently working on Deep learning based models for forecasting of solar wind and CME. Figure 2 shows the result of a deep learning model for solar wind prediction. The initial work is carried out using Convolutional Neural Network(CNN) for solar wind forecasting. We developed a CNN model from scratch to forecast solar wind speed from SDO/AIA images. Following figure shows the prediction of our CNN based model for slow and fast solar wind 4-days in advance. This forecasting scheme can predict both the fast and slow wind well with a RMSE of 76.3±1.87 kms-1 and an overall correlation coefficient of 0.57±0.02 for the year 2018, while significantly outperforming bench-mark models. The threat score for the model fares around 0.56. In 65% cases, the proposed model can accurately forecast the occurrence of HSEs. 

Figure 3 shows the regions identified by the CNN model responsible for slow and fast wind. This indicates the advantages of using a data science approach for mining new knowledge.