Aug 17 – 21, 2026
National Institute for Space Research, São José dos Campos, SP, Brazil
America/Sao_Paulo timezone

High-Cadence Solar Flare Dynamics via Deep Learning: Soft X-ray Spectral Analysis for for Aditya-L1/SoLEXS

Not scheduled
20m
Fernando de Mendonça - LIT (National Institute for Space Research, São José dos Campos, SP, Brazil)

Fernando de Mendonça - LIT

National Institute for Space Research, São José dos Campos, SP, Brazil

Av. dos Astronautas, 1758 - Jardim da Granja, São José dos Campos - SP, 12227-010
Oral Machine Learning in Space, Earth & Atmospheric Sciences Oral Contributions

Speaker

Abhilash Sarwade (U. R. Rao Satellite Centre, ISRO)

Description

Solar Low-Energy X-ray Spectrometer (SoLEXS) is a Sun-as-a-star payload onboard the Aditya-L1 mission designed to monitor solar coronal emissions and flare energetics. It has been in continuous operation at the L1 Lagrangian point for almost two years, capturing solar soft X-ray (SXR) spectra at a 1-second cadence. The stability of the instrument and its observing conditions has produced a large, homogenous, and high-cadence dataset ideal for machine learning analysis. However, analyzing this 1-second data presents a significant computational challenge. The study of rapid plasma heating during solar flares, a key science goal, is fundamentally limited by the trade-off between temporal resolution and signal-to-noise ratio (SNR). While key dynamic processes occur on shorter timescales, the low SNR of these short-integration spectra makes traditional forward-modeling techniques infeasible, as they require long integration times (e.g., 30-180s) for reliable fitting. This temporal averaging obscures the fast dynamics targeted for investigation.

A novel analysis pipeline is presented, designed to overcome this limitation and extract plasma parameters (e.g., temperature, elemental abundances) at the full 1-second cadence from SoLEXS spectra. The method employs a hybrid ‘Teacher-Student’ deep learning approach. First, a ‘Teacher’ model, consisting of a fully connected neural network, is trained on reliable, high-SNR time-integrated spectra, learning to rapidly emulate the results of traditional fitting. A second, recurrent model (LSTM) is then trained to learn the temporal dynamics, using sequences of noisy 1-second spectra to predict the parameters of the integrated interval. This allows the pipeline to leverage both the reliability of high-SNR data and the temporal information in the 1-second-cadence sequences, effectively ‘de-noising’ the data by learning how plasma parameters evolve smoothly in time.

This pipeline produces a 1-second-cadence time series of plasma parameters, capturing rapid variations previously lost in time-averaged analysis. This high-cadence output reveals new phenomena. Notably, less-intense flare events are observed that are typically missed by traditional, long-integration analysis. Furthermore, the 1-second temperature model output reveals evidence of quasi-periodic pulsations (QPPs) in the thermal emission. This technique opens a new observational window, enabling the quantitative study of fast heating and cooling processes in the flare plasma at the instrument's native cadence.

Author

Abhilash Sarwade (U. R. Rao Satellite Centre, ISRO)

Co-author

Dr K. Sankarasubramanian (U. R. Rao Satellite Centre, ISRO)

Presentation materials

There are no materials yet.