Abstract Details


Name: Amrit Roy
Affiliation: National Institute of Techonology Srinagar
Conference ID: TVS202510229
Title: EdgeDeploy-SolarNet: Compact Deep Learning for Real-Time Solar Flare Prediction
Authors and Co-Authors: Aayusha Singh
Abstract Type: Contributory Presentation
Abstract: Space-based solar observation missions need to predict solar activity in real time, but they have to work with very limited computing power and memory whereas, current deep learning models for solar flare prediction often have computational requirements that are not well-suited to the limited power, memory, and processing resources available on spacecraft and edge computing platforms. EdgeDeploy-SolarNet is introduced as a framework for developing computationally efficient neural networks designed for solar flare classification in resource-constrained environments. The framework makes use of CNN and LSTM architectures trained on GOES 15/16 X-ray flux time series along with SDO/HMI SHARP magnetic field parameters to predict M-class and X-class flares. To enhance efficiency, model compression is applied through post-training quantization, structured pruning, and knowledge distillation. Computational constraints similar to those of ARM-based processors are simulated to assess the feasibility of deployment. Performance evaluation shows that 8-bit quantization reduces model size by 75% with 92% of the baseline accuracy for 24-hour flare prediction. Pruned models provide a fivefold reduction in parameters with less than 5% loss in accuracy. The compressed models also achieve inference latencies suitable for real-time operation on simulated edge hardware. Feature analysis further suggests that the lightweight models preserve physically interpretable magnetic field signatures linked to flare-productive active regions. The proposed framework meets key needs for autonomous space weather prediction systems, where computational efficiency is essential. The results demonstrate that advanced machine learning models for solar flare prediction can be deployed on resource-constrained platforms, supporting improved space weather monitoring for future missions such as CubeSats, interplanetary probes, and distributed sensor networks.