AI-FLARES: Artificial Intelligence for the Analysis of Solar Flares Data

Michele Piana${}^{1,2}$, Federico Benvenuto${}^{1}$, Anna Maria Massone${}^{1}$, Cristina Campi${}^{1}$, Sabrina Guastavino${}^{1}$, Francesco Marchetti${}^{3}$, Paolo Massa${}^{4}$, Emma Perracchione${}^{5}$, Anna Volpara${}^{1}$

${}^{1}$ MIDA, Dipartimento di Matematica, Università di Genova, Genova, Italy
${}^{2}$ Osservatorio Astrofisico di Torino, Istituto Nazionale di Astrofisica, Pino Torinese, Italy
${}^{3}$ Dipartimento di Matematica “Tullio Levi Civita”, Università di Padova, Padova, Italy
${}^{4}$ Department of Physics & Astronomy, Western Kentucky University, Bowling Green, KY 42101, USA
${}^{5}$ Dipartimento di Scienze Matematiche “Giuseppe Luigi Lagrange”, Politecnico di Torino, Torino, Italy

Abstract

AI-FLARES (Artificial Intelligence for the Analysis of Solar Flares Data) is a research project funded by the Agenzia Spaziale Italiana and by the Istituto Nazionale di Astrofisica within the framework of the “Attività di Studio per la Comunità Scientifica Nazionale Sole, Sistema Solare ed Esopianeti” program. The topic addressed by this project was the development and use of computational methods for the analysis of remote sensing space data associated to solar flare emission. This paper overviews the main results obtained by the project, with specific focus on solar flare forecasting, reconstruction of morphologies of the flaring sources, and interpretation of acceleration mechanisms triggered by solar flares.

key words. Sun: flares – Data: EUV, magnetograms, hard X-rays – Methods: artificial intelligence

1 Introduction

Solar flares are the most explosive and energetic events that characterize the active Sun (Tandberg-Hanssen & Emslie, 1988). They may extend for more than 10,000 kilometers, release more than $10^{32}$ ergs in less than $100$ seconds, accelerate billion of tons of materials to more than one kilometer per hour, and generate electromagnetic radiation at all wavelengths. From a physical viewpoint, this source of radiation is characterized by very low resistance and very high inductance, so that the rising phase of a flare should last for an (impossibly) long time, which is in contrast with respect to empirical observations. This flare paradox, together with the fact that very little is known about the acceleration and energy release mechanisms within the flaring region, are the reasons why solar flares are still a hot topic in both experimental and theoretical solar plasma physics. Further, from a technological viewpoint, flares are the trigger of space weather (Moldwin, 2022), i.e., the physical and phenomenological state of natural space environments, which may notably impact the technological assets at earth.

Two general investigation issues concerning solar flares are:

1. Are the information contained in the magnetic fields constrained within solar active regions (ARs) and measured by means of magnetograms accurate enough to allow a reliable flare forecasting process?

2. Which physical mechanisms determine the acceleration of the electrons in coronal plasma, thus triggering the process that leads to the emission of high energy radiation via bremsstrahlung with the ions of the ambient plasma?

The “Artificial Intelligence for the analysis of solar FLARES data (AI-FLARES)” project, funded in 2019 by the Agenzia Spaziale Italiana and the Istituto Nazionale di Astrofisica under the “Attività di Studio per la comunità scientifica nazionale Sole, Sistema Solare ed Esopianeti” framework, recognized that these issues can be accomplished by means of two different approaches. From the one hand, magnetohydrodynamics (MHD) equations can be numerically reduced in order to simulate the flaring mechanisms. However, this approach is significantly hampered by the complexity of these partial differential equations and by the limited accuracy of the numerical approximation of their solutions. On the other hand, data-driven approaches can exploit the notable amount of solar, heliophysics, and space weather missions that are currently operating and that can provide an unprecedented amount of multi-modal measurements concerning essentially all possible manifestations of the active Sun. Given this available wealth of data, AI-FLARES focused on the formulation and implementation of computational methods for their interpretation, with applications to the forecasting and modelling of solar flares. Specifically, AI-FLARES developed computational methods for the prediction of the flaring emission and the identification of flare precursors, the reconstruction of flare morphologies for intense eruptive events, and the comprehension of the energy release and acceleration mechanisms for both thermal and non-thermal electrons. The project’s objectives were to

• Reconstruct the saturated EUV signal in the core region of images of flaring storms recorded by the Atmospheric Imaging Assembly on-board the Solar Dynamics Observatory (SDO/AIA) (Lemen et al., 2012).

• Provide an imaging-spectroscopy picture of the acceleration mechanisms at the base of solar high-energy emissions by exploiting visibilities recorded by both the Reuven Ramaty High Energy Solar Spectroscopy Imager (RHESSI) (Hurford et al., 2003) and the Spectrometer/Telescope Imaging X-rays (STIX) on-board Solar Orbiter (Krucker et al., 2020).

• Design flare forecasting processes and identify the most significant precursors of intense flares by applying machine learning and deep learning algorithms to the Helioseismic and Magnetic Imager on-board SDO (SDO/HMI) magnetograms (Scherrer et al., 2012).

The methodological inspiration of AI-FLARES relied on an extended interpretation of artificial intelligence, including supervised machine and deep learning, image processing, inverse problems and inverse diffraction theory. From a technological viewpoint, the outcome of this research effort has been a set of computational pipelines for the interpretation of flare-related physics that can be reached at https://github.com/theMIDAgroup/AI-FLARES.

The plan of the paper is as follows. Section 2 describes AI-FLARES results concerned with flare forecasting. Section 3 and Section 4 focus on image processing and reconstruction at EUV and hard X-ray wavelengths, respectively. Our conclusions are offered in Section 5.

2 AI-FLARES and solar flare forecasting from magnetograms

In the last decade a notable amount of scientific literature has been illustrating the results of the application of data-driven AI-based approaches to flare forecasting (Camporeale et al., 2018; Huang et al., 2018; Florios et al., 2018; Leka et al., 2019; Georgoulis et al., 2021; Ribeiro & Gradvohl, 2021; Nishizuka et al., 2021; Sinha et al., 2022). Most of these studies describes the action of either machine learning tools that process features extracted from magnetograms, or deep neural networks that directly take as input full disk images of solar active regions (ARs). On the one hand, a possible objective of feature-based machine learning is to identify which AR descriptors mostly impact the forecasting process. On the other hand, deep learning aims at improving the predictive power hidden within the space data by means of black-box approaches that take as input images or videos of AR magnetograms and provide as output a binary classification based on features that are automatically computed by the neural networks.

AI-FLARES addressed the first issue by means of sparsity-enhancing regularization methods applied to the SDO/HMI archive in specific but large time ranges (Benvenuto et al., 2018; Florios et al., 2018; Massone et al., 2018; Piana et al., 2018; Campi et al., 2019; Benvenuto et al., 2020; Cicogna et al., 2021). In a typical pipeline of analysis, the HMI magnetograms have been grouped into four subsets belonging to the four issuing times 00:00, 06:00, 12:00, and 18:00 UT. For each AR we used the $171$ features extracted by means of the algorithms implemented within the FLARECAST Horizon 2020 project (Georgoulis et al., 2021). For each subset, i.e., for each issuing time, we generated the training, validation, and test sets and, for each sample in the training set, the annotation was performed providing label ”1” to an event occurred within $24$ hours from the issuing time. By applying AI-FLARES machine learning algorithms to these data we were able to prove that (see Figure 1):

• Very few AR descriptors are really effective in the forecasting process and these descriptors are very robust, independently of the regularization method used and of other experimental aspects like the issuing times considered in the training set (Campi et al., 2019).

• The Ising energy seems to systematically play a notable predictive role, specifically in the case of the forecasting of particularly intense flaring storms (Benvenuto et al., 2020).

• The computation of innovative topological descriptors can represent a way to improve the skill scores associated to feature-based machine learning algorithms (Cicogna et al., 2021).

Figure 1: Feature-based machine learning for flare forecasting. Top left panel: very few descriptors (x-axis) are sufficient to achieve high values of the True Skill Statistic (TSS) score (y-axis) with two machine learning methods (hybrid LASSO, HLA and Random Forest, RF). Bottom panel: feature-ranking applied to machine learning outcomes show that the Ising energy significantly increases its rank when the AR producing the flaring event is included in the training set. Top right panel: the TSS notably increases when the topological feature introduced in Cicogna et al. (2021) is added as first descriptor (red solid line), with respect to the case wehm the feature is not used in the training set (blue dashed line).

The second AI-FLARES perspective on flare forecasting was the development of deep learning networks able to provide a notable prediction improvement in the difficult game of space weather prediction. In this context, the main result of AI-FLARES (see Figure 2) has been the implementation of a pipeline that, for the first time, utilizes videos of HMI frames as input data and that, again for the first time, accounts for an appropriate balancing of different data sample types in the training and validation phases of the neural network (Guastavino et al., 2022). We implemented the AI-FLARES Long-term Recurrent Convolutional Network (LRCN) and validated it against the Near Realtime Space Weather HMI Archive Patch (SHARP) data products associated with the line-of-sight components in the time range between 2012 September 14 and 2017 September 30. More specifically, we used 24-h-long videos made of 40 SHARP images of an AR, with 36 minutes cadence. These videos have been categorized as C-, M-, and X-class flares and also according to four different null-events classes. We generated a training set, a validation set and a test set, and we used data augmentation to increase their cardinality. We repeated this set generation process ten times in order to create ten random realizations of these three sets. The algorithm for set generation was inspired to two strategic principles: proportionality, i.e., use of the same rates of samples for each category; and parsimony, i.e., use of as few ARs as possible so that samples belonging to the same AR fall into the same set. As shown in Figure 2, the true positive rates provided by the deep network are significantly high (in particular, both M and X classes are more distinguishable from ARs associated with NO-class), and in all cases the standard deviations are nicely small.

As a final comment, we point out that AI-FLARES provided contributions also to the methodological field related to machine and deep learning research. In particular, during this project two theoretical ideas have been conceived and formulated. The first one is about the use of probabilistic score-oriented loss functions in the training phase for neural networks (Marchetti et al., 2022); the second one is about the use of value-weighted skill scores for the performance assessment of both machine and deep learning (Guastavino et al., 2022, 2023). These two methodological tools have been utilized in most networks designed for the flare forecasting approaches developed within AI-FLARES.

Figure 2: Deep learing for flare forecasting. Left panel: the AI-FLARES neural network is made of a Long Short-Term Memory (LSTM) network fed by the outcomes of several Convolutional Neural Networks (CNNs). Right panels: the rates of true positives and true negatives are significantly high thanks to the use of video-based deep learning.

3 AI-FLARES and EUV image de-saturation

EUV measurements recorded in correspondence with intense solar flares are almost systematically affected by saturation. After the launch of SDO/AIA, the desaturation of EUV images has become a big data issue, since AIA has been providing more the $10^{5}$ frames per year since February 2010 (Schwartz et al., 2014; Torre et al., 2015; 2015A&C....13..117S).

AI-FLARES contribution to image processing at the solar EUV regime is represented by the formulation and implementation of the Sparsity-Enhancing DESAT (SE-DESAT) method, a novel computational approach for the analysis of SDO/AIA saturated images able to recover the signal in the primary saturation region in a rapid fashion without using any other information but the one contained in the image itself (Guastavino et al., 2019). SE-DESAT is a modification of a previous algorithm developed in our group, named DESAT (Schwartz et al., 2015). As for DESAT, also in SE-DESAT the input data are represented by the diffraction fringes and therefore this is again an inverse diffraction algorithm. However, unlike for DESAT, this new approach realizes segmentation between the primary saturation region and blooming, background estimation and desaturation in the primary saturation region at the same time, without the need of any a priori information on the image background. Further, an adaptive version of SE-DESAT (adaptive SE-DESAT) introduced weights depending on the shape of the saturated region (Guastavino & Benvenuto, 2021). An example of how SE-DESAT performs is described in Figure 3, in the case of EUV images recorded by AIA on September 25 2011, in the $193$ $\AA$ bandwidth, at time point 03:33:31 UT.

Figure 3: Desaturation of EUV maps. Left panel: an intense solar flare saturates an extended region of an image recorded by SDO/AIA. Right panel: AI-FLARES desaturation algorithm is able to restore the signal in the core of the flaring source.

We point out that the availability of this desaturation pipeline has currently two important consequences. First, scientists working in the STIX community are used to integrate the hard X-ray information contained in the STIX data with the EUV information contained in the AIA data, and to this aim de-saturated AIA images are needed (Massa et al., 2022). Second, AI-FLARES people are currently involved in a NASA project for nowcasting solar flares starting from AIA information (Massa & Emslie, 2022). Once again, image desaturation will be a crucial pre-processing step for the realization of a prediction approach that will apply machine learning to de-saturated EUV maps.

4 AI-FLARES and hard X-ray imaging spectroscopy

AI-FLARES results concerned with hard X-ray imaging spectroscopy (Piana et al., 2022) have been of three kinds. We formulated the mathematical model for the image formation process of STIX visibilities and contributed to the calibration process for all thirty STIX collimators devoted to imaging (Massa et al., 2019; Krucker et al., 2020; Massa et al., 2023). We then developed several image reconstruction methods able to represent in the image space the information contained in the hard X-ray visibilities recorded by either RHESSI or STIX. Specifically, we have formulated, implemented and validated (see Figure 4):

• A maximum entropy method, in which the solution is constrained to have positive entries and total flux equal to an a priori estimate (Massa et al., 2020).

• An interpolation/extrapolation method based on feature augmentation and on the use of Variably Scaled Kernels (Perracchione et al., 2021a, b) that allows the implementation of an automated version of CLEAN deconvolution (Perracchione et al., 2023).

• A parametric imaging method that works for both a partial information on the visibility set, i.e., when just the visibility amplitudes are provided by the instrument (Massa et al., 2021) and when visibilities are recorded by fully calibrated collimators (Volpara et al., 2022) (the algorithm relies on Particle Swarm Optimization).

Finally, and probably more importantly, as part of the RHESSI legacy (Piana et al., 2007; Prato et al., 2009), we formulated and implemented a regularization method that is able to reconstruct maps whose pixel content is proportional to the average flux of the electrons accelerated along the magnetic field lines from the corona down to the chromosphere. The nicest aspect of this methodological approach (see Figure 5) is that it is able to provide electron maps that are constrained to vary in a smooth way along the spectral direction and that can be projected back to the photon domain to produce photon maps that are in turn regularized across energy. This approach may represent an important step to a full interpretation of STIX data within the framework of imaging spectroscopy, and may provide a crucial tool for the understanding of electron acceleration mechanisms during solar flares (Volpara et al., 2023).

Figure 4: Image reconstruction from hard X-ray visibilities. A constrained maximum entropy method, a forward-fit algorithm based on Particle Swarm Optimization (PSO), and an automated version of CLEAN deconvolution metod provided the reconstructions in the left, middle, and right panels, respectively.

Figure 5: Regularized imaging spectroscopy from hard X-ray visibilities. These panels represent electron flux maps at different electron energies (regularization introduced a smoothing constraint across the energy direction.)

5 Conclusions

This paper has reviewed the main results obtained within the framework of the AI-FLARES project. These results involve the whole workflow concerning the interpretation of solar flare data, starting from the prediction of flaring events using magnetograms, through the restoration of morphological aspects of the flaring sources in saturated high-resolution EUV maps, to the realization of an innovative imaging spectroscopy approach in the case of hard X-ray visibilities.

The legacy of AI-FLARES in the current activity of our research group is two-fold. On the one hand, we are developing neural networks for space weather forecasting that encode physical information in the loss function utilized for the training (Guastavino et al., 2023). On the other hand, we are studying how regularized electron maps can be used to obtain quantitative information about the effectiveness of the electron acceleration mechanisms triggered by magnetic reconnection high in the solar corona (Volpara et al., 2023).

6 Introduction

Imaging spectroscopy in solar hard X-rays, i.e., the construction of spatio-spectral cubes describing a flaring source in the energy range between a few and a few hundred keV, was the central objective of the NASA Reuven Ramaty High-Energy Solar Spectroscopic Imager (RHESSI) mission (2002SoPh..210....3L). Because of its rotating modulation collimator design (1965ApOpt...4..143O), RHESSI native measurements were in the form of two-dimensional spatial Fourier components of the incoming photon flux, termed visibilities (2002SoPh..210...61H), and this inherent aspect of the data drove the formulation, implementation, and application of several innovative approaches to imaging spectroscopy for RHESSI. Notable among these approaches was one (Piana et al., 2007; Prato et al., 2009) that combined regularized spectral inversion and Fourier-based image reconstruction methods to produce images of the flux of bremsstrahlung-producing electrons at different energies or, equivalently, electron flux spectra at different points in the image. The Piana et al. (2007) approach exploits the linearity of both spatial Fourier transform and photon $\rightarrow$ electron spectral inversion operations to reverse the conventional order of spatial and spectral inversions. Instead of performing a spatial Fourier transform on the observed count visibilities to produce count images, using these images to produce count spectra at different locations, and then spectrally inverting the count spectra at each location, the method reverses the order of spatial and spectral processing. It takes the observed count visibilities and, in conjunction with an assumed bremsstrahlung cross-section (e.g., 1959RvMP...31..920K), performs a regularized spectral inversion procedure (all in the spatial-frequency domain) to arrive at the visibilities associated with the flux spectrum of the electrons responsible for the bremsstrahlung emission. Due to the regularization (pianabook) inherent in their construction, these electron visibilities vary smoothly from one electron energy channel to the next, and it must be noted that this property is not shared by the native count visibilities. Count visibilities are constructed from independent bundles of detected counts in each prescribed energy range and, since each bundle of counts has its own independent level of statistical noise, the measured count number can vary erratically from one count energy channel to the next. This feature is exacerbated by the (ill-posed) nature of the spectral inversion problem (1986ipag.book.....C), resulting in inferred electron spectra that can possess large, erratic (and doubtless unphysical) variations with energy, and are hence of little use in addressing physical issues related to electron acceleration and transport. Spatial inversion, via a discrete Fourier transform, of the electron visibilities yields images of the electron flux at different electron energies $E$, and, because of the smooth energy variation of the electron visibilities from which they are constructed, these images also vary smoothly with electron energy. This allows a more reliable (and indeed more sophisticated) physical investigation into the variation of the electron spectrum from point to point within the flare. Among the scientific results provided by this approach using RHESSI data, we mention the determination of the acceleration region size in flaring loops (2012A&A...543A..53G), of the number of particles within the flare acceleration region (2012ApJ...755...32G), of the values of the emission filling factor and the specific acceleration rate (2013ApJ...766...28G), and of the parameters associated with stochastic acceleration models (2010ApJ...712L.131P). In the present era, visibility-based observations of solar flare hard X-ray emission are obtained with the Spectrometer/Telescope for Imaging X-rays (krucker2020spectrometer), one of the six remote-sensing instruments on board the ESA cluster on the Solar Orbiter mission (2020A&A...642A...1M). In contrast with the rotating modulation collimator approach of RHESSI, which generated visibilities by means of a rather involved data-stacking process (2002SoPh..210...61H), STIX computes visibilities in a completely static way, at a fixed set of points in the spatial- (or more correctly, angular-) frequency domain, using the properties of the Moiré patterns produced by pairs of grids with slightly different geometry (Massa et al., 2023). Although, for a given flare, the number of measured RHESSI visibilities could in principle have been quite large, in practice the signal in many cases was not sufficient to guarantee a reliable visibility measurement, particularly at higher energies, with their lower count rates. Furthermore, the spatial resolution of the instrument was compromised by blurring effects introduced by the data stacking process. By contrast, each one of the $30$ STIX visibilities is provided by a unique STIX collimator, so that all visibilities are observed at all time points and at all energies, without any blurring effects. The first objective of the present work is to provide an implementation of the regularized imaging spectroscopy method that was introduced for RHESSI, but which is now optimized for the STIX framework discussed above. We have also chosen, in contrast to the algorithm developed for RHESSI, to design an algorithm that can be applied to non-uniformly-distributed energy channels. By considering a specific case study, we illustrate the STIX high-level research products that can be generated by means of regularized imaging spectroscopy; we discuss how the reliability of such products can be validated; and we provide a quantitative estimate of the higher accuracy of STIX electron flux spectral images compared to the ones obtained from RHESSI. Finally, we use this regularized imaging spectroscopy approach to show that, with STIX observations, the spectral behavior of accelerated electrons in solar flares can be tracked from the corona to the chromosphere, with associated implications for the nature of electron transport mechanisms in solar flares. The plan of the paper is as follows. Sect. 7 summarizes the regularized imaging spectroscopy approach in the case of visibility-based telescopes, while Sect. 8 applies this approach to experimental measurements recorded by STIX. Sect. 9 discusses preliminary results obtained with the regularized imaging spectroscopy technique in the case of a solar flare recorded on May 8, 2021 (SOL2021-05-08T18). Our conclusions are offered in Sect. 10.

7 Electron flux spectral images

We here follow Appendix A of Piana et al. (2007) in reviewing the various concepts, and related quantities, that are involved in the construction of images (maps) of the electron flux at different energies from STIX observations of count visibilities.

7.1 Underlying concepts

Define a Cartesian coordinate system $(x,y,z)$ such that $(x,y)$ (each in arcsecond units) represents a location in the image plane and $z$ represents distance (measured in cm) along the line of sight. Let the local density of target particles within a source of line-of-sight depth $\ell(x,y)$ (cm) be $n_{\rm target}(x,y,z)$ (cm${}^{-3}$) and let the electron flux spectrum, differential in energy $E$, at the point $(x,y,z)$ in the source be $F(x,y,z;E)$ (electrons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$). We formally define the mean electron flux spectrum ${\overline{F}}(x,y;E)$ (electrons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$) by (pianabook)

$${\overline{F}}(x,y;E) = \frac{\int_{z=0}^{\ell(x,y)}n_{\rm target}(x,y,z)\,F(x,y,z;E)\,dz}{\int_{z=0}^{\ell(x,y)}n_{\rm target}(x,y,z)\,dz}$$ (1)

$$\equiv \frac{\int_{z=0}^{\ell(x,y)}n_{\rm target}(x,y,z)\,F(x,y,z;E)\,dz}{{\overline{n}}_{\rm target}(x,y)\,\ell(x,y)}\,\,\,\,,$$ (2)

where ${\overline{n}}_{\rm target}(x,y)$ is the density of target protons, averaged along the line of sight distance $\ell(x,y)$. The corresponding photon spectrum image $I(x,y;\epsilon)$ (photons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$ arcsec${}^{-2}$) produced by bremsstrahlung encounters of these electrons with ambient protons is (2003ApJ...595L.115B)

$$I(x,y;\epsilon) =\frac{a^{2}}{4\pi R^{2}}\cdot$$ (4)

$$\cdot\int^{\infty}_{E=\epsilon}\int_{z=0}^{\ell(x,y)}n_{\rm target}(x,y,z)\,F(x,y,z;E)\,Q(\epsilon,E)\,dE\,dz$$

$$= \frac{a^{2}}{4\pi R^{2}}{\overline{n}}_{\rm target}(x,y)\,\ell(x,y)\int^{\infty}_{E=\epsilon}{\overline{F}}(x,y;E)\,Q(\epsilon,E)\,dE\,\,\,\,,$$ (5)

where $R$ (cm) is the distance from the source to the instrument, $Q(\epsilon,E)$ (cm${}^{2}$ keV${}^{-1}$) is the bremsstrahlung cross-section, differential in photon energy $\epsilon$, for photon emission at energy $\epsilon$, and $a=7.25\times 10^{7}\,{\widetilde{R}}$ cm arcsec${}^{-1}$ is the conversion factor from arcseconds to cm, ${\widetilde{R}}$ being the distance from the source to the observer in astronomical units (au). Eq. (4) may be written

$$I(x,y;\epsilon)=\frac{1}{4\pi R^{2}}\,\int^{\infty}_{E=\epsilon}\mathcal{F}(x,y;E)\,Q(\epsilon,E)\,dE\,\,\,,$$ (6)

where we have defined the electron flux spectral image (electrons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$ arcsec${}^{-2}$)

$$\mathcal{F}(x,y;E)=a^{2}\,\,{\overline{n}}_{\rm target}(x,y)\,\ell(x,y)\,{\overline{F}}(x,y;E)\,\,\,.$$ (7)

We can treat the set of electron flux spectral images either as maps of the electron flux at a prescribed electron energy $E$, or as a set of mean electron flux spectra at each point $(x,y)$ in the image plane. Given that the STIX instrument on Solar Orbiter observes the Sun from different distances at different times in its heliocentric orbit, a brief digression into the role of the source-observer distance $R$ is pertinent, and also yields some valuable insights. For a given target density $n_{\rm target}(x,y,z)$ and electron flux $F(x,y,z;E)$ profile, the observed photon spectrum image, i.e., the hard X-ray emission from a region of prescribed angular extent, is, according to Eq. (4), independent of the source-observer distance $R$. At a smaller (say) value of $R$, a given solid angle extent (in arcseconds${}^{2}$) corresponds to a smaller area on the source plane. The cm${}^{2}$-to-arcsecond${}^{2}$ conversion factor $a^{2}$ is accordingly smaller, leading to the multiplicative factor $a^{2}/4\pi R^{2}$ being independent of $R$. Physically, a smaller number of photons are emitted from the reduced physical area corresponding to a square arcsecond of angular extent. However, this factor is exactly cancelled by the reduced source-observer distance in the inverse-square law relating the number of emitted photons and the number that impact a distant detector. Thus, while a given flare event will result in more recorded counts for an instrument located at ${\widetilde{R}}<1$ than for an identical instrument at $1$ au, this is because the angular extent of the source is larger, and not because the photon spectrum image $I(x,y;\epsilon)$ is any brighter.

7.2 Determining electron flux spectral images from STIX observations

The STIX instrument (krucker2020spectrometer) on the Solar Orbiter mission (2020A&A...642A...1M) does not directly image hard X-rays. Rather, a set of $30$ segmented detectors, each located behind a pair of absorbing grids with different geometrical properties (Massa et al., 2023), record count measurements, in each of four detector segments, at different energies. This information is then combined to obtain the values of $30$ two-dimensional spatial Fourier components, or visibilities, of the source structure at the count energy in question. Each visibility corresponds to a specific angular frequency $(u,\varv)$, where $u$ and $\varv$ are in units of arcsec${}^{-1}$ (see Fig. 10 of Massa et al., 2023). The relationship between a count visibility $V(u,\varv;q)$ (counts cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$), recorded at energy $q$, and the photon images $I(x,y;\epsilon)$ (photons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$ arcsec${}^{-2}$) at different photon energy values $\epsilon$ is given by (Piana et al., 2007)

$$V(u,\varv;q)=\int_{X}\,\int_{Y}\,\int_{\epsilon=q}^{\infty}\,D(q,\epsilon)\,I(x,y;\epsilon)\,e^{2\pi i(ux+\varv y)}\,d\epsilon\,dx\,dy~{},$$ (8)

where the spatial integrals extend over the entire field of view of the instrument, and $D(q,\epsilon)$ (keV${}^{-1}$) is the value of the detector response matrix corresponding to the probability per unit count energy of a count with energy $q$ being produced by a photon of energy $\epsilon$. We note that while the distance from the source to an instrument does not change the magnitude of the photon image for a given event (since, as discussed in Sect. 7.1, the photon image is defined as the emissivity per arcsec${}^{2}$ on the sky, and at a smaller source-observer distance a specified angular size corresponds to a smaller physical distance on the source), the field of view of STIX is sufficiently large that observed visibilities measure the emission from the entire Sun, and hence the magnitude of the visibilities $V(u,\varv;q)$ do increase as the distance from the Sun to the instrument decreases. Combining Eqs. (6) and (8) leads to the mathematical expression that links the electron flux spectral images $\mathcal{F}(x,y;E)$ to the STIX count visibilities $V(u,\varv;q)$, namely (cf. Eq. (A6) of Piana et al., 2007)

$$\begin{split}&V(u,\varv;q)=\frac{1}{4\pi R^{2}}\cdot\\ &\cdot\int_{X}\,\int_{Y}\,\int_{\epsilon=q}^{\infty}\,\int_{E=\epsilon}^{\infty}\mathcal{F}(x,y;E)\,D(q,\epsilon)\,Q(\epsilon,E)\,e^{2\pi i(ux+\varv y)}\,dE\,d\epsilon\,dx\,dy\,\,\,.\end{split}$$ (9)

Reversing the order of integration over $\epsilon$ and $E$ gives

$$V(u,\varv;q) =\frac{1}{4\pi R^{2}}\cdot$$ (11)

$$\cdot\int_{X}\,\int_{Y}\,\int_{E=q}^{\infty}\int_{\epsilon=q}^{E}\,\mathcal{F}(x,y;E)\,D(q,\epsilon)\,Q(\epsilon,E)\,e^{2\pi i(ux+\varv y)}\,d\epsilon\,dE\,dx\,dy$$

$$= \frac{1}{4\pi R^{2}}\int_{X}\,\int_{Y}\int_{E=q}^{\infty}\mathcal{F}(x,y;E)\,K(q,E)\,e^{2\pi i(ux+\varv y)}\,dE\,dx\,dy\,\,\,,$$ (12)

where we have defined the count cross section

$$K(q,E)=\int_{\epsilon=q}^{E}D(q,\epsilon)\,Q(\epsilon,E)\,d\epsilon\,\,\,,$$ (13)

with units of cm${}^{2}$ keV${}^{-1}$, differential in count energy $q$. This quantity combines the bremsstrahlung cross-section for production of a photon of energy $\epsilon$ by an electron of energy $E$ with the probability (per unit count energy) that such a photon will be recorded as a count with an energy $q$, to give the overall cross-section, differential in count energy, for production of a count of energy $q$ by an electron of energy $E$. Finally, for each $(u,\varv)$ we introduce the electron flux visibility spectrum (electrons cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$)

$$W(u,\varv;E)=\int_{X}\,\int_{Y}\,\mathcal{F}(x,y;E)\,e^{2\pi i(ux+\varv y)}\,dx\,dy\,\,\,,$$ (14)

the Fourier transforms of the electron flux spectral image at the measured spatial frequencies $\{(u_{i},\varv_{i})\}_{i=1}^{30}$ and electron energy $E$. Then, changing the order of integration over space and electron energy in the last equality of Eq. (11), we finally derive the relationship between the observed differential count visibility spectrum $V(u,\varv;q)$ and the underlying electron flux visibility spectrum:

$$V(u,\varv;q)=\frac{1}{4\pi R^{2}}\,\int_{q}^{\infty}W(u,\varv;E)\,K(q,E)\,dE\,\,\,.$$ (15)

Eq. (15) is the fundamental result presented in Piana et al. (2007), and its similarity to Eq. (4) essentially shows that the spectral relationship between count and electron quantities in the spatial domain carries through to the spatial frequency domain. Electron flux visibility spectra $W(u,\varv;E)$, which contain information on the spatial (more accurately, angular on the plane of the sky) distribution of the electron flux spectral images are not, of course, directly measured by STIX, but they can be retrieved from the observed count visibility spectra $V(u,\varv;q)$ by inverting Eq. (15) via any number of established inversion techniques (pianabook). Specifically, for a given angular frequency $(u,\varv)$, the (complex) values of $W(u,\varv;E)$ at different energies can be retrieved by applying a well-tested Tikhonov regularization approach (1994A&A...288..949P; 2003ApJ...595L.127P) to both the real and the imaginary parts of Eq. (15). We refer the reader to Piana et al. (2007) and to Prato et al. (2009) for more details about the implemented inversion method. The electron flux spectral images for different energies $\mathcal{F}(x,y;E)$ can be constructed from the set of electron flux visibility spectra $\{W(u_{i},\varv_{i};E)\}_{i=1}^{30}$ by applying standard image reconstruction methods (e.g., pianabook). We remind the reader that as a result of the regularized spectral inversion procedure used to generate the electron flux visibilities $W(u,\varv;E)$ these visibilities, and hence the electron flux spectral images $\mathcal{F}(x,y;E)$ that are constructed from them, vary smoothly with energy $E$, unlike the count visibilities $V(u,\varv;q)$ on which they are based. As we shall see, this feature of the electron flux spectral images permits a more impactful scientific analysis than is possible from the study (e.g., 2008ApJ...673..576X) of count-based images.

Figure 6: The SOL2022-09-29T14 event. Top row, left panel: position of the event on the solar disk; right panel: normalized STIX time profiles corresponding to the energy ranges 9-12 keV and 22-76 keV. The reported time is relative to Solar Orbiter. Bottom row: count images of the event at different energies as reconstructed by MEM$\_$GE.

As noted earlier, this visibility-based imaging spectroscopy methodology was originally developed in the RHESSI framework. Inspired by that approach, here we developed an algorithm that is tailored to the use of STIX visibilities. From a technical viewpoint, there are two main differences between the two methods. First, for RHESSI, at some $(u,\varv)$ points the count visibilities $V(u,\varv;q)$ often suffered gaps at specific $q$ values, as a consequence of the sub-optimal data stacking process that characterizes the underlying rotating modulation collimator (RMC) imaging concept; instead, for STIX there is, by design, a one-to-one correspondence between sampled visibilities and collimators, so that count visibility spectra are always complete. Second, the implementation of the algorithm realized for RHESSI constrained the count visibility spectra to be uniformly sampled along the count energy direction. Instead, the STIX algorithm considers non-uniformly-spaced count energy channels by introducing appropriate weights in the quadrature formula for the discretization of Eq. (15) and by accounting for such weights in the regularized inversion formula.

8 Application to STIX visibilities

In order to assess the electron flux spectral image methodology using STIX visibilities, we considered a case study represented by the near-limb (as viewed from Solar Orbiter) flare observed by STIX on September 29, 2022 (SOL2022-09-29T14). Although this flare occurred on the far side of the Sun as seen from Earth, from STIX observations the estimated GOES class is $\sim$M1, with an uncertainty range from C7 up to M2. The top row panels of Fig. 6 show the position of the source on the solar disk and the (normalized) STIX time profiles corresponding to the energy ranges 9-12 keV and 22-76 keV. The two bottom rows contain the count images reconstructed by applying MEM$\_$GE, a constrained maximum entropy algorithm (Massa_2020) to the visibility bag corresponding to the time interval between 14:36:19 and 14:37:19 UT (relative to Solar Orbiter). As we now discuss, the regularized imaging spectroscopy approach allows users to obtain research products of higher level than simply count images; they include regularized electron flux spectral images, regularized count images, and spatially-resolved regularized count and electron flux spectra.

8.1 High-level research products

Figure 7: Electron flux images in different energy channels for the SOL2022-09-29T14 event, as reconstructed by MEM$\_$GE.

The algorithm described in the previous section, using the isotropic form of the bremsstrahlung cross-section in 1959RvMP...31..920K, provides regularized electron flux visibility spectra for each $(u,\varv)$ point sampled by STIX. These data can be re-ordered by collecting together all values of the regularized solutions $W(u,\varv;E)$ of Eq. (15) corresponding to the same electron energy $E$. The application of the MEM$\_$GE algorithm to each of these electron flux visibilities provides images of the electron flux at different energies, as shown in Fig. 7. As a consequence of the regularization process involved in constructing the electron flux visibilities used to produce the electron flux spectral images, these images necessarily vary smoothly with electron energy. Furthermore, the algorithm provides visibilities at electron energies that are larger than the highest count energy observed in the count images of Fig. 6; this is due to the fact that counts registered in a specific energy channel result from photons that are in turn produced via bremsstrahlung by electrons with all energies higher than that of the photon, so that counts at a given energy provide information on electrons of all higher energies. A second high-level research product can be obtained by substituting the (regularized) electron visibility spectra into the bremsstrahlung-like equation (15) to produce count visibility spectra that are also regularized. Applying an inverse spatial Fourier transform to these produces the regularized count images of Fig. 8. Unlike the raw count-based images, these regularized images have the desirable property of varying smoothly with (count) energy. Further, the count cross section $K(q,E)$ in Eq. (15) can be replaced with the bremsstrahlung cross section $Q(\epsilon,E)$ to obtain regularized photon visibilities and hence, via application of a spatial inverse Fourier transform, regularized photon images. For these products, the instrumental dependency is removed.

Figure 8: Top panels: Regularized count images in different energy channels for the SOL2022-09-29T14 event, as reconstructed by MEM$\_$GE. Bottom left panel: selected locations are indicated with colored crosses. Middle column: spectra at different locations obtained from the regularized count images. Right column: pixel-wise spectra obtained from the regularized electron flux spectral images.

Perhaps the most sophisticated high-level research products provided by regularized imaging spectroscopy are the ones contained in Figs. 8 and 9. Taking into account the limited spatial resolution of STIX, we average the values of the regularized count images and electron flux spectral images over areal boxes of $5$ arcsec $\times\,5$ arcsec size and then construct regularized count and electron spectra for each such $25$ arcsec${}^{2}$ area. The spectrum at each such location can then be analyzed, e.g., by fitting with parametric models. Thanks to the smoothing effects of regularization, this fitting process is significantly more reliable than the one applied to non-regularized spectra; indeed, the regularized spectra suffer less spurious oscillations and are characterized by uncertainties that are significantly smaller than the ones associated with non-regularized spectra. Below are a few comments on the results presented in the third column of the bottom panels of Fig. 8.

• For the northern footpoint (indicated by a red plus sign in the left panel of Fig. 8), the corresponding panel of Fig. 8 clearly indicates both thermal (below 20 keV) and non-thermal (above 20 keV) components of the electron flux spectrum, with the transition between these regimes evidenced by the spectral flattening above 20 keV.

• The area highlighted by a green plus sign corresponds to an area shifted by 10 arcsec towards the South relative to the northern source. Its spectrum shows a very similar behavior; however, as the image angular resolution is $\sim$14 arcsec, the spectra for these two regions are not completely independent of each other.

• The spectrum associated with the southern source (indicated with an orange plus sign in the left panel of Fig. 8) is harder to interpret. As it can be seen in Fig. 7, this source is visible in the electron flux images only above 16 keV and, therefore, the spectral points below this energy are not reliably determined.

• The spectrum corresponding to the area in between the northern and southern sources is indicated with a blue plus sign. At this location, the electron flux spectrum is similar to the other sources indicated; however, considerations of spatial resolution and dynamic range demand that caution must be taken in the interpretation of this spectrum. While the STIX dynamic range can be as large as $\sim$20:1 for a single strong source (see Sect. 4.3 of krucker2020spectrometer), limited data statistics and complex source morphology reduce this to a value that is typically $\sim$10:1 for images such as those presented in Figs. 6 through 8. Only clearly identifiable sources within the dynamic range of the STIX instrument in the considered energy range can be reliably be used for scientific analyses, and the electron flux images of Fig. 7 do not show a clearly identifiable source at this location. Thus, the inferred electron spectrum in this region is likely a result of the influence of the other nearby sources in in the image.

Similar considerations apply to the regularized count spectra shown in the second column of the bottom panels of Fig. 8.

8.2 Validation

Figure 9: Spatially resolved count spectra for the SOL2022-09-29T14 event. Top left panel: Selected pixels are indicated with colored crosses. Other panels show the corresponding pixel-wise spectra obtained from count images (solid line) and regularized count images (dashed line) as a function of count energy. The pixels selected in the top left panel and their respective spectra are indicated with the same color. The plots are logarithmically scaled on the $y$-axis. The ($3\sigma$) error bars on the count spectra are in pink, while the error bars on the regularized count spectra are in black.

Event	OSPEX	electron flux spectral images	count images	regularized count images
SOL2021-05-08T18	$\gamma=5.28\pm 0.18$	$\delta=4.53\pm 0.07$	$\gamma=4.94\pm 0.24$	$\gamma=5.50\pm 0.04$
SOL2021-08-26T23	$\gamma=5.65\pm 0.12$	$\delta=4.59\pm 0.11$	$\gamma=5.27\pm 0.17$	$\gamma=5.57\pm 0.06$
SOL2022-01-20T05	$\gamma=6.51\pm 0.26$	$\delta=5.07\pm 0.03$	$\gamma=6.25\pm 0.14$	$\gamma=6.36\pm 0.01$
SOL2022-08-28T16	$\gamma=6.74\pm 0.18$	$\delta=4.97\pm 0.04$	$\gamma=6.88\pm 0.24$	$\gamma=6.81\pm 0.03$
SOL2022-09-29T14	$\gamma=4.54\pm 0.05$	$\delta=3.68\pm 0.01$	$\gamma=4.24\pm 0.11$	$\gamma=4.42\pm 0.01$

Table 1: First column: event date. Second column: photon power-law index $\gamma$ obtained by fitting spatially integrated spectra with OSPEX. Third column: electron power-law index $\delta$ obtained by fitting spatially integrated electron flux spectral images. Fourth and fifth columns: photon power-law index $\gamma$ obtained by fitting spatially integrated count images and spatially integrated regularized count images, respectively, assuming a unit detector response matrix. The uncertainties in columns 3, 4, and 5 are determined by repeating the computation for 10 random realizations of the input visibilities.

In order to assess the validity and reliability of the regularized imaging spectroscopy method, we studied the impact that regularization has on both spatially-resolved and spatially-integrated spectra. Using the same SOL2022-09-29T14 event as the case study, Fig. 9 considers the five positions in the flare highlighted in the top left panel and, for each one, compares the local count spectra provided by the original images in the bottom rows of Fig. 6 with the ones provided by the regularized count images in Fig. 8. The five plots clearly show that while regularization does not modify the overall shape of the local count spectra, it does increase their stability, in particular for locations in the flaring source where the signal-to-noise ratio is smaller. For thin-target hard X-ray emission (appropriate for calculating the count spectrum image $J(x,y;q)$ (counts cm${}^{-2}$ s${}^{-1}$ keV${}^{-1}$ arcsec${}^{-2}$) produced by a mean electron flux spectrum $F(x,y;E)$ at a given position $(x,y)$), use of relatively simple (e.g., Kramers, Bethe-Heitler) forms of the bremsstrahlung cross-section, and assuming a diagonal Detector Response Matrix $D(q,\epsilon)$ with constant entries, lead to a simple approximate relation between the power-law spectral indices $\gamma$ and $\delta$ of the count spectrum $J(q)$ and the electron flux spectrum $F(E)$, viz. $\gamma=\delta+1$ (Eq. (1.24) of pianabook). By comparison, for thick-target emission in a target with a (collisional) mean free path that is proportional to $E^{2}$, the target-averaged electron flux spectrum is correspondingly two powers harder than the injected electron spectrum (Eq. (1.19) of pianabook), so that $\gamma$ is related to the power-law index $\delta_{\rm inj}$ of the injected spectrum by the oft-cited relation $\gamma=(\delta_{\rm inj}-2)+1$, i.e., $\gamma=\delta_{\rm inj}-1$ (Eq. (1.25) of pianabook). Table 1 compares the spectral indices provided by spatially integrating the count images, the electron flux spectral images, and the regularized count images with the ones provided by the standard spectroscopy procedure implemented in OSPEX, in the case of five events characterized by similar morphologies. As for both the spatially-integrated count images and regularized count images, the corresponding spectral index has been fitted assuming a unit detector response matrix. This approximation does not significantly affect the outcomes of this analysis since, over the bulk of the energy range used to fit the spectral index, the STIX detector response matrix is approximately diagonal with entries that vary over a modest range from $\sim$0.36 to $\sim$0.42), insufficient to cause the spectral index of the count spectrum to differ significantly from that of the incident photon spectrum. The entries in the table show that 1) the count spectral indices generated by the regularized count images reproduce well the ones provided by OSPEX; and 2) the approximate thin-target relation $\gamma=\delta+1$ are both well satisfied by the regularized electron flux and count spectra.

8.3 STIX vs. RHESSI

RHESSI and STIX are both visibility-based instruments, which in general offer the possibility of applying Fourier-based methods for image reconstruction and, more specifically, to use the approach described in Sect. 7 to create regularized electron flux and photon spectrum images. However, while the ways RHESSI and STIX generate visibilities both involve occulting grids, they are intrinsically very different. Each of the RHESSI collimators samples a circle in the $(u,v)$-plane via rotation of the spacecraft. In the case of the STIX instrument, each sub-collimator is designed to measure a single Fourier component. As a consequence, while RHESSI in principle can measure hundreds of visibilities, some of them have an insufficient signal-to-noise ratio, and all of them suffer a blurring effect related to the data stacking process; by contrast, STIX provides $30$ complex visibility values that are observed at all times, with better statistics and without any blurring effects. The higher quality of STIX imaging spectroscopy with respect to RHESSI is demonstrated in Table 2, where we consider three events observed by STIX and three events observed by RHESSI, all with comparable statistic and morphologies. We reconstructed both the regularized electron flux and regularized count images at different energies using CLEAN (Clean; Perracchione et al., 2023). For each event and each energy channel we identified the maximum in the residuals map and in the CLEANed map as defined in 2009ApJ...698.2131D and computed the ratio; then for each event, we averaged this ratio across all energies and computed the corresponding standard deviations. The results in Table 2 show that, for all cases, this ratio is systematically and significantly higher for RHESSI than for STIX, which explains the higher statistical reliability of STIX visibilities when used to obtain the regularized electron spectrum images and count images.

Event	counts	electrons
SOL2022-09-29T14 (STIX)	0.061 $\pm$ 0.007	0.069 $\pm$ 0.019
SOL2002-02-20T11 (RHESSI)	0.114 $\pm$ 0.027	0.171 $\pm$ 0.067
SOL2023-01-11T01 (STIX)	0.058 $\pm$ 0.012	0.065 $\pm$ 0.035
SOL2003-12-02T22 (RHESSI)	0.124 $\pm$ 0.025	0.292 $\pm$ 0.225
SOL2022-11-11T01 (STIX)	0.070 $\pm$ 0.013	0.084 $\pm$ 0.032
SOL2002-02-20T11 (RHESSI)	0.128 $\pm$ 0.029	0.167 $\pm$ 0.038

Table 2: Analysis of CLEAN reconstructions for regularized counts and electron flux spectral images obtained using RHESSI and STIX visibilities. The second and third columns show the ratio of the residual and CLEANed maps peaks for the regularized count and electron flux spectral images, respectively, averaged across count and electron energies.

9 Analysis of the SOL2021-05-08T18 event

For most flares the non-thermal hard X-ray emission from the chromospheric foot-points outshines the much fainter coronal sources at higher energies $\gtrsim 20$ keV, making an analysis of the variation of the electron spectrum along the flare loop very difficult, especially for visibility-based imaging systems such as STIX that necessarily see every source in the field of view. On the other hand, coronal thick-target flares (see 2004ApJ...603L.117V) have a higher-than-normal density in the corona, so that very few electrons have sufficient energy to reach the chromospheric footpoints. Consequently, most of the non-thermal emission is from the relatively dense corona, and the resulting absence of footpoint emission makes it feasible to observe how the shape of the (line-of-sight-integrated) nonthermal mean electron flux spectrum $\int F(x,y,z;E)\,dz$ varies along the projection of the flaring loop on the plane of the sky. In such cases the inferred variation of the local electron spectrum $F(E,s)$, at distance $s(x,y)$ from the injection point along the direction defined by the magnetic field lines linking the coronal and chromospheric region of the flare, allows electron transport models to be tested (e.g., 2001ApJ...557..921E; 2017ApJ...851...78A). In the following we show preliminary electron spectra from a flare observed by STIX and characterized by a prominent non-thermal component from the corona, similar to the prototypical coronal thick-target flares from 2004ApJ...603L.117V. This analysis takes advantage of the enhanced imaging dynamic range provided by regularized imaging spectroscopy. On May 2021 two active regions generated a series of C- and M-class flares observed by STIX, GOES, and SDO/AIA (Massa et al., 2022). We concentrate here on the event recorded on May 8 in the time interval (relative to Solar Orbiter) between 18:29 and 18:32 UT (SOL2021-05-08T18). The top left panel of Fig. 10 shows the (normalized) light-curves for both GOES and two STIX energy ranges, for the time interval considered for this analysis. The top right panel shows the AIA image, re-projected to match the viewing angle from Solar Orbiter/STIX and, overlaid, the contours inferred through application of the MEM$\_$GE algorithm to STIX electron visibilities integrated between 25 and 28 keV. Since the flare ribbons are extended, projection effects could have an effect on the interpretation of the actual 3-D structure of the event. Nevertheless, a loop-like structure connecting the UV flare ribbons, and visible over a wide range of electron energies, is evident in the STIX images. The bottom left panel of Fig. 10 shows the electron flux spectral image in the energy channel between 25 and 28 keV, computed using the regularized imaging spectroscopy code; superimposed are crosses representing eight selected locations within the loop structure. In the bottom middle and right panels, we show the spatially resolved electron flux spectra at each of these locations, with the middle and the right panels showing the spectra corresponding to the locations in the left and right branches of the loop, respectively. The spectra are obtained by averaging the pixel values inside a box of 5 $\times$ 5 pixels around the locations highlighted in the bottom–left panel of Fig. 10. The uncertainties shown by the error bars are obtained by means of the confidence strip approach (1994A&A...288..949P). The input STIX count visibilities are perturbed 25 times with Gaussian noise with a standard deviation equal to the experimental uncertainty. Then, the electron visibilities and the corresponding electron maps are retrieved from each perturbed data and finally the standard deviation of the spectral intensities obtained from the different electron map realizations is computed.

Figure 10: The SOL2021-05-08T18 event. Top left panel: normalized GOES and STIX time profiles. Top right panel: X-ray contour plots, provided by the MEM$\_$GE algorithm, overlaid on the AIA map, which has been re-projected to match the viewing angle from STIX. The $10$, $30$, $50$, $70$, and $90$% contour levels of the reconstructed non-thermal X-ray emissions (25 - 28 keV) are plotted in magenta. Bottom left panel: electron flux map in the energy channel 25-28 keV, with seven locations identified in the loop. Bottom middle panel: regularized electron flux spectra at four locations in the left side of the flaring loop with corresponding uncertainties. Bottom right panel: regularized electron flux spectra at five locations in the right side of the flaring loop, with corresponding uncertainties. The spectra are obtained by considering areas of $5\times 5$ arc seconds around each of the selected points and averaging the corresponding spectra. The spectra clearly show two components: a thermal component below $\sim$20 keV and a non-thermal component at higher energies.

The spectra clearly show two components: a thermal component below $\sim$20 keV and a non-thermal component above $\sim$20 keV, with a clear minimum, or “dip,” in the overall spectrum around 20 keV. It has been shown (2005SoPh..232...63K) that such dips can be an artefact of neglecting to subtract the albedo component (that results from scattering of primary source photons from the photosphere; 1978ApJ...219..705B) from the observed hard X-ray spectrum. However, such artefacts are most pronounced in sources with fairly hard ($\gamma\simeq 2$) spectra, which have a proportionately larger number of higher energy photons that backscatter into the energy range being observed; therefore, given the much steeper spectrum ($\gamma\simeq 5.5$; Table 1) associated with this event, the effects of neglecting albedo photons is unlikely to be the reason for the observed dip. Further, Fig. 4 of 2002SoPh..210..407A shows that the albedo correction is most pronounced at energies substantially larger than 20 keV; even for flat photon spectra the correction to the inferred electron flux spectrum at 20 keV is only some 10%, not enough to create an artificial dip of size comparable to those in Fig. 10. We also note that the albedo source typically has a significantly larger spatial extent than the primary source (2002SoPh..210..273S) and therefore is unlikely to make a substantial contribution to the electron flux spectrum in a relatively localized region, such as those considered for computing the spectra shown in Fig. 10. We therefore conclude that the observed dips are indicative of a real minimum in the primary source electron flux spectrum. Further, the evolving shapes of the nonthermal components show a phenomenological behavior that is plausibly consistent with the effects of electron dynamics within the loop: as the distance from the loop top (the purported location of the electron acceleration region) increases, electrons of progressively higher energies are stopped, slowed, or thermalized through interaction with the electrons in the ambient plasma. Thus the value of the electron flux decreases at all energies, the local minimum at the interface between the thermal and nonthermal components becomes more pronounced, and the energy corresponding to the maximum in the nonthermal component shows a tendency to increase. We shall study this behavior, and its consistency with a collisional model (e.g., 1972SoPh...26..441B; 1978ApJ...224..241E) for the electron dynamics, more quantitatively in a subsequent work.

10 Comments and Conclusions

The first objective of this study was to show that the regularized imaging spectroscopy approach introduced for the analysis of RHESSI visibilities is also feasible in the STIX framework. Further, we have shown that the peculiar one-to-one correspondence between visibilities and collimators realized by the Moiré pattern technique utilized in STIX allows the generation of research products (namely, regularized electron spectral images and count images, and regularized spatially resolved electron flux and count spectra) that have significantly better quality than those obtained with the rotating modulation collimator approach of RHESSI. Thanks to this higher reliability, we have been able to show, for the first time, how the spectrum of the accelerated electron flux varies along the path from the corona to the dense chromospheric regions at the loop footpoints, presumably through interaction of the accelerated electrons with the ambient plasma. One aspect that may hamper the reliability of the outcomes of this approach is the limited dynamic range provided by the STIX detectors. Indeed, the spectra in Fig. 10 show ranges of intensities which, for all energy channels, are almost always close to this 10:1 limit, across all locations selected within the loop. Nevertheless, it is clear that this sort of analysis cannot be performed for all flares, since frequently the dynamic range is a limiting factor in constructing accurate spectra. For example, such an analysis will be likely impossible to perform for flares with prominent footpoint emission, since footpoint emission is typically much brighter than the emission from the flare legs. The best candidates would be flares with a prominent coronal source such as coronal thick-target sources (2004ApJ...603L.117V) or flares with occulted footpoints. For the latter a case, a second observatory that sees the entire flare site (see, e.g., 2019RAA....19..167K) could provide additional information on the spectral shape of the footpoints. Regularized imaging spectroscopy with STIX visibilities is therefore probably the first technique that could allow the testing of energy loss rates (for example, either Coulomb or non-Coulomb collisional losses; see Sect. 4 of 2001ApJ...557..921E) for non-thermal electron transport. Given that such loss rates in general depend on the ambient density, future work will be devoted to investigate whether it is possible to determine the ambient density through the dependence of the electron spectrum on position, and thus separate the mean electron flux and the ambient electron density factors in the bremsstrahlung equation (4). The thus-determined ratio of the accelerated electron population to the ambient density in the flaring region constitutes an important constraint on the effectiveness of electron acceleration mechanism(s) in solar flares.