A machine learning approach for computing solar flare locations in X-rays on-board Solar Orbiter/STIX

This repository contains the dataset and the code developed for addressing the problem of estimating solar flare locations from data recorded by the Coarse Flare Locator (CFL) on board the Spectrometer/Telescope for Imaging X-rays (STIX) with a neural network. The work is described in a paper that will appear soon on arxiv.

Dataset

The dataset folder contains the training, validation and test set files. Each file contains the following information:

• nn_data: array of dimension N_samples x 9 containing the CFL data that are given as input to the MLP
• stx_flare_loc: array of dimension N_samples x 2 containing the X and Y coordinates of the flare locations derived from STIX imaging information
• cfl_loc: array of dimension N_samples x 2 containing the X and Y coordinates of the flare locations estimated by the CFL algorithm
• start_time: array of dimension N_samples containing the start observation time of each example
• end_time: array of dimension N_samples containing the end observation time of each example
• tot_counts: array of dimension N_samples containing the total number of counts registered by the STIX imaging detectors for each example
• tot_counts_cfl: array of dimension N_samples containing the total number of counts registered by the STIX CFL detector for each example
• sidelobes_ratio: array of dimension N_samples containing an estimate of the reliability of the flare location derived from STIX imaging (ratio between the peak value and the sidelobes' maximum value in the full-disk Back Projection image)
• file_uid: array of dimension N_samples containing the UIDs of the STIX FITS files from which the data are extracted

Model definition and training

The train_nn.ipynb notebook shows how the Multi-Layer Perceptron (MLP) is defined and trained. Auxiliary functions used for defining the model and for normaliziong the dataset contained in aux_code.py.

Quantized model predictions on the test set

The test_quantized_nn.ipynb notebook shows how the neural network inputs, weights and biased are quantized in such a way that all the operations in the forward pass are performed in 16-bit fixed-point integer arithmetic. The operations performed by the quantized neural network are implemented in nn_model. A comparison between the flare location estimates provided by quantized MLP and those obtained by the CFL algorithm on the test set examples is shown in the notebook.