| Name: | Linn Abraham |
| Affiliation: | Inter-University Centre for Astronomy and Astrophysics |
| Conference ID : | ASI2024_344 |
| Title : | Interpretable Deep Learning for Solar Flare predictions |
| Authors : | Linn Abraham, Vishal Upendran, Durgesh Tripathi, Ninan Sajeeth Philip, Nandita Srivastava, A. Ramaprakash, Sreejith Padinhatteeri |
| Authors Affiliation: | 1. Linn Abraham ( Inter-University Centre for Astronomy and Astrophysics, Pune 411 007, India,
Puducherry Technological University, Puducherry, 605 014, India)
2. Vishal Upendran ( Bay Area environmental research institute, moffet field, CA, USA, Lockheed Martin Solar and Astrophysics Laboratory, Palo Alto, CA, USA)
3. Durgesh Tripathi, A. Ramaprakash ( Inter-University Centre for Astronomy and Astrophysics, Pune 411 007, India)
4. Ninan Sajeeth Philip (Artificial Intelligence Research and Intelligent Systems, Kerala 689544, India)
5. Nandita Srivastava (Udaipur Solar Observatory, Physical Research Laboratory, Udaipur 313001, India)
6. Sreejith Padinhatteeri (Manipal Academy of Higher Education, Manipal 576104, India)
|
| Mode of Presentation: | Poster |
| Abstract Category : | Sun, Solar System, Exoplanets, and Astrobiology |
| Abstract : | Within the past decade or more several Machine learning (ML) based models have been developed for the prediction of solar flares predominantly making use of the magnetogram data. Shallow, interpretable, ML models have been historically employed, operating on numerous derived features from magnetograms, with the algorithms showing nearly similar performance upon optimization. This similarity in performance may result from the application of the same features derived from photospheric magnetograms. Progress on using the original magnetogram or coronal imaging measurements has been minimal due to data complexity and suitable computational models. Deep learning models provide us with the avenue to consume multiwavelength data to perform flare forecasting. In this work, we seek to generate an understanding of intensity measurement on the Sun responsible for flares. For this purpose, we develop a Deep Learning (DL) model and train it to classify AIA Active Region data cubes into flaring and non-flaring classes. We then employ interpretable AI tools like Grad CAM, Shapley values, and Integrated gradients to understand the specific features most important for performing such a classification. This helps us open up the black box DL model, giving us insights into the physics of flare trigger mechanisms. |
