Identifying Typical Mg ii Flare Spectra Using Machine Learning

The data for this study were taken by NASA's small explorer satellite IRIS, which was launched in 2013 and has since observed more than 400 flares according to a manually kept list on its mission website. IRIS can take high quality slit-jaw images (SJI) of the solar atmosphere with a maximum field of view of 175 × 175 arcsec² in four different passbands, covering a range of heights from the photosphere to the transition region. It is also equipped with a spectrograph that can run simultaneously at high spatial (0.33–0.4 arcsec), spectral (0.056 Å in the NUV), and temporal (2s) resolutions. Various observing modes can be selected, from sit-and-stare, where the spectrograph slit remains stationary with respect to the Sun, to rasters with varying numbers of steps and slit orientations.

We use an unsupervised clustering algorithm known as k-means to identify the different Mg ii k profiles that occur during a flare. Clustering algorithms partition similar observations into groups. If each observation consists of two features, they can be mapped as points on the Cartesian plane. k-means uses additional points called centroids to group the observations. The centroids are placed on the plane and each observation is assigned to one of the centroids with a straight line. The k-means objective is to find the assignments and position of centroids that minimize the sum of all the squared Euclidian distances. This quantity is known as the "within cluster distance" ${ \mathcal L }$ and is given by

$$\begin{eqnarray}&&{ \mathcal L }=\displaystyle \sum _{i=1}^{n}\displaystyle \sum _{j=1}^{k}{\delta }_{{c}_{i},j}| | {x}_{i}-{\mu }_{j}| {| }^{2}.\end{eqnarray} \tag{ 1 }$$

Here, ${\delta }_{{c}_{i},j}$ is the Kronecker delta and c_i is a label assigning each observation ${x}_{i}\in {{\mathbb{R}}}^{{d}_{x}}$ to one of the k-groups, so that

$$\begin{eqnarray}{\delta }_{{c}_{i},j}=\left\{\begin{array}{ll}1, & \mathrm{if}\ {x}_{i}\,\mathrm{belongs}\ \mathrm{to}\ \mathrm{group}\,j,\\ 0, & \mathrm{if}\ {x}_{i}\,\mathrm{does}\ \mathrm{not}\ \mathrm{belong}\ \mathrm{to}\ \mathrm{group}\,j,\end{array}\right.\end{eqnarray} \tag{ 2 }$$

while ${\mu }_{j}\in {{\mathbb{R}}}^{{d}_{x}}$ is the centroid of that group. To initialize the algorithm, the centroids are positioned in the locations of k randomly selected data points. The k-means algorithm then minimizes the within cluster distance by iterating through a two step procedure known as coordinate descent. First, the data is partitioned into groups by labeling each observation by the centroid it appears closest to:

$$\begin{eqnarray}&&{c}_{i}=\mathop{\arg \,\min }\limits_{1\leqslant j\leqslant k}| | {x}_{i}-{\mu }_{j}| {| }^{2}.\end{eqnarray} \tag{ 3 }$$

Second, the centroids are moved to the mean position of each group according to

$$\begin{eqnarray}&&{\mu }_{j}=\displaystyle \frac{1}{{n}_{j}}\displaystyle \sum _{i=1}^{n}{\delta }_{{c}_{i},j}\,{x}_{i},\end{eqnarray} \tag{ 4 }$$

where n_j is the number of all observations in group j. Because the centroids have shifted, new labels must be assigned to all the observations in accordance with step one. This process continues until the centroids converge. The final clustering depends on the initialization of the centroids. It is customary to repeat the clustering with a number of different centroid initializations and select the clustering with the lowest cost ${ \mathcal L }$. k-means was chosen for its simplicity, scalability, and linear time complexity (Hartigan & Wong 1979). The above algorithm is easily adapted to the purposes of grouping line profiles. Instead of two features being mapped to a two-dimensional Cartesian plane, we composed each profile out of 216 λ points and mapped them to a 216-dimensional space (i.e., d_x = 216). k-means requires the number of clusters k to be chosen by hand. There are many methods, such as the "elbow technique" and silhouette analysis (Rousseeuw 1987), that indicate the natural number of partitions of a data set; however, this number in many cases is left to the discretion of a professional and should not be automated. In conclusion, k-means generates k-groups, with each group containing a number of spectra that share similar features. These groups each have a representative spectral profile, which is the mean profile of that group called the centroid.

We experimented with a number of alternative distance metrics, most of which returned similar if not identical results to the Euclidean distance, while other distance measures found clusters that were hard to reconcile with any physical interpretation. Additionally, when updating the centroids, one could select the median or a true representative point instead of the arithmetic mean. This could make the algorithm less susceptible to the effects of outliers at the price of being more computationally expensive. However, such effects are negligible if there are a sufficient number of normal data points. Furthermore, we acknowledge that there may exist a transformation and metric pair that is more suited to the clustering of spectral data, but did not wish to interject any biases into our data set.

Figure 1 shows the result of k-means applied to a single raster of the M6.5 class flare on 2015 June 6. The bottom panel shows 6 of the 53 centroids found by our adapted k-means algorithm explained in Section 2.4. The color-coded positions of the spectra for this raster have been overlaid onto the SJI. The Mg ii k profiles emerging from each location are assigned to one of the 53 groups based on which centroid they most closely resemble.

We analyzed 26 M- and 7 X-class flares with observations in the Mg ii spectral window and slits positioned directly over the flaring region. The details of each flare can be found in Table 1, which includes a variety of observational modes.

Before using the k-means algorithm, the data were prepared in several ways. The Mg ii spectra were selected over a time interval 15 minutes before to 15 minutes after each flare, if available. The start and end times were determined from the GOES flare database. The spectra were then cropped to a 5 Å window that contained both the Mg ii k and subordinate lines consisting of the two strongest red wing $3{{\rm{d}}}^{2}{\rm{D}}\to 3{{\rm{p}}}^{2}{\rm{P}}$ transitions with vacuum wavelengths at 2798.75 and 2798.82 Å. This step not only reduces computational demands but also helps negate dimensionality problems common among machine learning algorithms with distance-based metrics. The profiles were then interpolated using a spectral sampling of 0.025 Å/pixel, resulting in all Mg ii k-line profiles having a total of 216 λ points. Since the line intensities in data numbers (DN) can vary over two orders of magnitude, each profile was divided by its maximum intensity. This ensures that the classification is only based on the shape of the profile and not on its intensity. Data loss from incomplete spectra was automatically handled during the cropping phase. To visualize the results, we projected the assigned spectra onto the 1400 Å slit-jaw image, which is sensitive to plasma at chromospheric- and transition-like temperatures; however, if this filter was not used for that particular observation, the 2796 Å slit-jaw image, which sees the lower chromosphere was substituted.

The k-means algorithm was applied to Mg ii spectra collected from the subset of 4 M- and 4 X-class flares with the number of groups set to k = 80. Flaring groups were generated by manually selecting only the spectra that appeared over the flaring regions. The algorithm was repeated 10 times with different centroid initializations, and the clustering with the lowest cost ${ \mathcal L }$ was selected (see Section 2.2). If two or more centroids were similar, only a single centroid was retained.

Because every profile has to be assigned to a group, it is necessary to collect a number of non-flaring centroids to filter out the non-flaring spectra. Following the same procedure above, we generated quiet Sun and sunspot groups by manually selecting only the spectra that appeared over quiet Sun and sunspot regions within the eight selected flares. Centroids associated with small energetic events due to flux emergence were generated by running k-means over rising flux regions. This completed the unsupervised part of the algorithm.

For the remaining flares, the centroids from the clustering were passed to another algorithm (one-nearest neighbor classifier), where each profile was labeled by the centroid they appeared nearest to. Once the profiles of every flare were assigned to their groups, the variances of the flaring groups were monitored. If the variance was too high, an additional flaring group was introduced manually, with the proviso that the new group consistently corresponded with some feature on the SJI projections (see Figure 1). This procedure is demonstrated in Figure 2, and was continued until we were satisfied that all interesting flare groups were included. The final 53 groups can be seen in Figure 3. Our method of clustering deviates from the original k-means algorithm in that we manually merge and split groups based on the supervision of both the variance and the SJI projections. It is unclear whether the same results could be achieved simply by increasing the initial number of centroids in the original k-means algorithm. We refer to this new clustering method as "supervised hierarchical k-means," or SHK for short.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The 53 profile types fall into four main classes: Quiet Sun, sunspot, flux emergence, and flaring centroids. Figure 4 shows examples of each of the four classes. The quiet Sun profiles have characteristic central reversals, well defined blue (2kv) and red (2kr) peaks and line wings that follow the temperature structure of the lower atmosphere. Flaring profiles on the other hand can be extremely broad and are often observed without central reversals and with the subordinate lines in emission. Similarly, sunspot profiles often have a single peak, but can be distinguished from flaring profiles based on their narrow width ∼0.25 Å and lack of subordinate line emission. Flux emergence profiles share many features of quiet Sun and flaring profiles. They have raised wings and are broad; however, there is no subordinate line emission.

Zoom In Zoom Out Reset image size

We investigate if all flares share common profiles, which would indicate that the physics of the lower solar atmosphere may be similar in all flares.

The single peaked profiles can be divided into two different types. The first is represented by centroid 0 and has small subordinate line emission and a FWHM of 0.5 Å. The second represented by centroid 4 and 5 has a narrower convex shape with large subordinate line emission and a FWHM between 0.29 and 0.49 Å. Broader profiles such as those belonging to groups 1, 2, and 3 are rarer, have less predictable behavior, and often appear with small central reversals. In Figure 5, we have plotted the average variances of groups 0, 4, 5, and 9 for each of the 33 flares. The single peaked profiles belonging to group 0 appear in every flare with a total average variance of 0.52, and are always located over the ribbon or in regions of small brightenings. These profiles are not exclusive to solar flares and can be stimulated by an assortment of sub-flare energetic events, such as local heating from small scale reconnections over flux emergence regions. We tested their prevalence by analyzing eight non-flaring energetic events, and one true quiet Sun observation. They did not appear in the quiet Sun observation but were seen in 7/8 of the energetic events (see Table 2). The profiles therefore link both high and low energetic solar activity with temporal occurrences before, after, and during the flare, and can be identified as a universal flaring profile.

Zoom In Zoom Out Reset image size

Single peaked profiles from group 4 and 5 with large triplet emission occur in the wake of the ribbon front and can have long characteristic life times of ∼1 hr (e.g., flare 31 in our list). The large triplet emission associated with these groups may be due to the heating of both the upper and lower atmosphere by non-thermal electrons (Pereira et al. 2015). In every flare, the frequency of these profiles is strongly correlated with the GOES 1–8 Å channel.

Here we investigate which profiles occur at ribbon fronts, where accelerated electrons are thought to reach lower atmospheric layers. We find that profiles belonging to groups 11 and 12 occur at the leading edge of fast-moving flare ribbons; however, there are also a number of false positives generated by upflowing material. These profiles are similar to the ribbon-front profiles but lack subordinate line emission. Profiles from groups 11 and 12 seem to be less exaggerated versions of the rarer profiles in group 52, which also appear on the ribbon front with broader widths, deeper central reversals, and larger subordinate line emissions. In Figure 6, we have plotted four flares during their impulsive phase. Panel A contains the upper ribbon of a two-ribbon flare and has profiles assigned to groups 11 and 12 at three different positions, each of which are following the direction of the ribbon, upwards for the highest point and downwards for the other two points. Similar behavior can be noted in panels B, C, and D. In each case, profiles from groups 11 and 12 follow the leading edge of the ribbon. Once the ribbon passes, profiles from groups 4 and 5 appear over the heated regions where the NUV emission is enhanced. When the ribbon starts to progress more slowly, the single peaked profiles from group 0 take their place, as can be seen in the online video.

Profiles belonging to group 52 have been observed by Rubio da Costa & Kleint (2017; flare 28 of our list) to occur co-spatially and temporally with hard X-ray emission in the RHESSI 32–100 keV energy band. We found supporting evidence for the alignment of these profiles with hard X-ray signatures in the M6.5 flare observed by both IRIS and RHESSI on 2015 June 22 (flare 22 of our list). As seen in Figure 7, one RHESSI hard X-ray source coincides with the cyan markers, which indicate the positions of profiles from group 52. The contours were drawn using the "clean" reconstruction algorithm from a time integration equal to the cadence of the IRIS raster. The roll angle of RHESSI was not modified. The slanted black lines show the locations of the IRIS raster. Unfortunately, RHESSI was in earth's shadow during the impulsive phase of the flare; therefore, we cannot be sure if the upper part of the ribbon had hard X-ray signatures previously, which may explain the black-colored profiles in that location and also the few cyan profiles seen in the video at earlier times. Black ribbon-front profiles could still be related to X-ray emission if the X-ray emission is at least 10 times fainter than that of the main source, making them invisible due to RHESSI's limited dynamic range. In the online videos, profiles from group 52 can be seen at the front of the fast-moving bottom ribbon between times T17:53:37 and T17:57:00, but not exactly at the time shown in the figure. This could have several explanations: The profiles may have coincided with hard X-rays at these times, but it cannot be verified due to RHESSI's orbit. The hard X-rays at 18:04 could be too weak to trigger the cyan profiles, or alternatively, hard X-rays do not necessarily trigger cyan-type profiles. They could also be hidden in this observation: post flare loops with large amounts of downflowing material can be seen in the IRIS video and faintly in this figure. The viewing angle means that IRIS observes the lower ribbon through the loops. This may explain why instead of profiles from group 52 lining the lower ribbon, we see triangular profiles with large redshifts. In summary, cyan profiles seem to occur near hard X-ray contours, but there are exceptions and a future statistical study is desirable.

In Figure 8 we have plotted the frequency of occurrence of the three ribbon-front IRIS profiles for eight flares. The GOES curve and derivative in the 1–8 Å channel have been included for each flare to outline the impulsive phase. The three panels of flares 28, 21, and 29 contain the ribbon centroids as well as 20 overplotted profiles from the flare in that panel. Profiles belonging to group 52 occur during the impulsive phase and cluster around the maximum of the GOES derivative, in contrast to groups 11 and 12, which are more dispersed in relation to the GOES derivative, and may be the result of lower energy electron bombardment appearing to follow soft X-ray signatures. Profiles assigned to group 52 have FWHMs of ∼1.5 Å and characteristic blueshifted reversals at 2796.27 ± 0.04 Å, whose wavelength calibration we have verified with quiet Sun profiles that were centered at 2796.34 Å. Furthermore, these profiles may have large non-thermal contributions to their line width. Since IRIS has a negligible instrumental broadening, the non-thermal velocities can be calculated using the formula from Milligan (2011) given by

$$\begin{eqnarray}&&{\xi }_{\mathrm{nth}}={\left[{\left(\displaystyle \frac{\mathrm{FWHM}}{{\lambda }_{0}}\displaystyle \frac{c}{2\sqrt{2\mathrm{log}2}}\right)}^{2}-\displaystyle \frac{2{k}_{B}T}{m}\right]}^{1/2},\end{eqnarray} \tag{ 5 }$$

where the second component is the squared Doppler velocity and the line formation temperature T was taken as 10^4.4 K. Assuming the profiles are resolved, we find an upper limit of 95 km s⁻¹ for the non-thermal velocities, neglecting both pressure and opacity broadening. The velocities would be diminished if the profiles were unresolved. Figure 9 shows the typical line shapes of ribbon-front IRIS profiles. The cyan profile represents the average profile from group 52 that appears in 42% of our flares. It is possible that IRIS misses the hard X-ray locations for some flares. The black profile is an average of every profile belonging to group 11 and 12 and occurs jointly in 100% of the observations. Note the striking similarity between the two averaged profiles. Both have central reversals at precisely the same wavelength, and emissions from both the strong far red wing and its partner line, forming a dimple. This leads us to believe that the ribbon profiles place a real physical constraint on the non-thermal velocity fields generated by the electron beam.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

The intensity ratios of the Mg ii h&k lines can be used as a diagnostic for the optical depth. A ratio of 2:1 of k:h indicates that the lines are formed in optically thin conditions, and a ratio of 1:1 indicates optical thickness. A derivation of these ratios is given in the Appendix.

In addition to the h&k lines, there exists a companion of triplet lines from transitions between the 3p²P and 3d²D states. These subordinate lines are located on the blue and red wings of the k-line and have vacuum wavelengths of 2791.60, 2798.75, and 2798.82 Å. The oscillation strength of the blue subordinate line (not visible in our window) is twice as weak as the far red line. From here on out, the strongest subordinate line at 2798.82 Å will be denoted by s.

Early observations by Lemaire et al. (1984) using NASA's Orbiting Solar Observatory recorded k/h ratios taken before and during a flare to be in the range 0.9–1.5. Quiet Sun ratios were measured by Kohl & Parkinson (1976) in the range 1.14–1.46, while more recent high resolution IRIS observations by Kerr et al. (2015) found time averaged quiet Sun ratios of 1.204 ± 0.010 and flaring ratios in the range 1.07–1.19. The SHK algorithm allows us to perform a large scale multi-flare study of the k/h ratios using the groups to partition flaring and non-flaring profiles. Additionally, we monitor the trend of the k/s ratios.

IRIS's sensitivity slowly degrades over time; therefore, the spectral data must be recalibrated for each observation. Before calculating the line ratios, we took into account the change in effective area and converted the measured intensities I_m(DN/s) into physical units ${I}_{\mathrm{phys}}(\mathrm{erg}\,{{\rm{s}}}^{-1}\,{\mathrm{cm}}^{-2}\,{\mathrm{sr}}^{-1}\,{\mathring{\rm A} }^{-1})$; see Kleint et al. (2016) for details. The ratios were then calculated by dividing the integrated intensities of the h&k lines taken over a 2 Å window. For a window exceeding 2 Å, the asymmetries in the wings, which contain a number of blended lines, result in inaccuracies.

Figure 10 shows the evolution of both the k/h and k/s line ratios. For times t < 0, the ratios of profiles belonging to quiet Sun groups were used, while profiles from flaring groups were used for times t ≥ 0. This division is marked with a black vertical line across all panels. The raw data appear as points, with each point representing the average measurements of ratios occurring within the specified minute window. In order to clearly show the trends, the data points of the k/h and k/s ratios for t ≥ 0 have been fitted with a third order polynomial applied over a running widow of width 9. The numerical scales of the k/h-line ratios for each flare are given on the right-hand side of each panel. It is clear that both the k/h and k/s ratios decrease during each flare, indicating an increase in opacity (due to enhanced electron densities) and enhanced subordinate line emission, respectively.

Zoom In Zoom Out Reset image size

In Figure 11, we partitioned the ratios into flaring and non-flaring ratios based on a few representative quiet Sun and flaring groups. The average k/h ratios for flaring profiles across all 33 flares were found to be 1.16 ± 0.02 in comparison to the quiet Sun values of 1.26 ± 0.03. The large variance in the ratios is a natural consequence of partitioning the profiles into groups. The results agree with the current ratio calculations of Kerr et al. (2015) based on flare number 15 of our list.

Zoom In Zoom Out Reset image size

The intensities in absolute units span a range of 10⁵–10⁸ $(\mathrm{erg}\,{{\rm{s}}}^{-1}\,{\mathrm{cm}}^{-2}\,{\mathrm{sr}}^{-1}\,{\mathring{\rm A} }^{-1})$. In Figure 12, we plotted the high intensity portion of ratios of all 33 flares, and color-coded ratios from profiles belonging to group 0, (11 + 12), and 52. The ratios remain far away from the optically thin y = 2x line, and in general, higher intensity profiles occur closer to flare maximum and correspond to larger opacities. These results are not surprising since both emission and opacity depend on electron density, which based on half-width measurements of high Balmer lines with τ₀ < 1 can reach values of n_e ∼ 4 × 10¹³ cm⁻³ at flare maximum, and vary over two orders of magnitude during a flare (Fritzová-Švestková & Švestka 1967). We conclude that the Mg ii lines not only remain optically thick during a flare, but appear closer to the 1:1 ratio than in the quiet Sun.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

In this section, we review the common features of flare profiles and compare these features to those of the quiet Sun. We refer to the known mechanisms of quiet Sun profile formation, and discuss the likely physical process responsible for the observed differences.

The quiet Sun profiles are observed almost completely at their τ = 1 value, and have a formation height that extends over the entire chromosphere, with the left (k1v) and right (k1r) minimums being formed in the lower chromosphere, left (k2v) and right (k2r) peaks in the middle chromosphere, and line core photons coming from the upper chromosphere, just below the transition region. The wings contain contributions from the photosphere, and because they are formed under LTE conditions at τ = 1, they follow the source function and consequently the lower atmospheric temperature structure. This results in raised wings as one samples lower and lower heights of the photosphere. The wings are most noticeable in profiles belonging to groups 44 and 45 in Figure 3. The raised wings become "flattened" for emissions over sunspots or during flares. For sunspots, the temperature is about ∼1500 K lower than the quiet Sun due to the stifling of convective energy by large kilogauss magnetic fields (e.g., Carlsson et al. 2015), which results in a diminished source function. For flares, the flattening could possibly be explained by a downward condensation layer that forms an optically thick barrier between the photosphere and the rest of the atmosphere (Kowalski et al. 2017). The resulting emergent intensities would then only contain contributions from the chromosphere.

In addition to the flattened wings, we found single peaked profiles to be prevalent in every flare and discuss their potential origin here. Once again, under quiet Sun conditions, the Mg ii h&k-line profiles have a characteristic central reversal. This reversal is common in strong chromospheric lines and is due to the source and Plank function decoupling at heights comparable to the core formation height. Jefferies & Thomas (1960) found a frequency independent approximation of the source function that holds for upper chromospheric heights, given by

$$\begin{eqnarray}&&S=\displaystyle \frac{\int {J}_{v}{\phi }_{v}{dv}+\epsilon {B}_{v}({T}_{e})}{1+\epsilon },\end{eqnarray} \tag{ 6 }$$

with mean intensity J_v, absorption profile ϕ_v, Planck function B_v(T_e), and photon destruction probability approximated as the ratio of the collisional and spontaneous de-excitation coefficients C/A. At heights close to the core formation height, the source function decreases on account of A being several orders of magnitude larger than C.

As explained by Leenaarts et al. (2013) and Leenaarts et al. (2012), the h&k-line core photons are produced in the thin upper chromosphere where large photon mean free paths and low photon destruction probabilities prevail. As such, the radiation field is horizontally smoothed and the Eddington approximation holds. Consequently, the source function along with the emergent intensity decreases with height, resulting in an observed central reversal. We find that during a flare, the central reversals are commonly seen in emission. Equation (6) demonstrates that an increase in electron density and temperature could result in a source function that continues to increase with height. A recent parameter study by Rubio da Costa & Kleint (2017) found that central emissions were indeed stimulated by either a large temperature spike in the upper chromosphere or an increase in electron density at the same location. In both cases, the coupling of the source and Planck function persists to heights above the core formation height, allowing the entire profile to form under LTE. The study demonstrated that unresolved up- and downflows can also lead to single peaked profiles formed under non-LTE conditions. In this case, the complex velocity fields containing both condensation and evaporation patterns can produce photons that fill the central reversal. It is unclear which of the three mechanisms are responsible for the single peaked flare profiles, which may result from an interplay between all three processes.

The ribbon would provide the most complementary conditions for single peak production, with enhanced electron densities, temperatures, and Doppler velocities, and therefore their prevalence in flares may be explained.

Recent observations of the ribbon front have shown that NUV spectra resulting from the non-thermal electron beam may differ from other spectra within the field of view. Xu et al. (2016) observed a negative flare front in He i 10830 Å using the Goode Solar Telescope at Big Bear Solar Observatory. The ribbon front produced the broadest Mg ii line profiles in the field of view, with a FWHM of 1 Å. Tei et al. (2018) observed the leading edge of a C-class flare kernel on 2014 November 11 with IRIS. They observed Mg ii profiles with intensity enhancements in the blue wing and smaller blue (h2v) peaks in comparison to the red (h2r) peaks of the h-line. The apparent blueshifts lasted 9–48 s, with speeds of 10.1 ± 2.6 km s⁻¹ and were followed by strong redshifts up to 51 km s⁻¹. They proposed a simple model where non-thermal energetic electrons would heat a deep region of the atmosphere, which would then expand and carry cool chromospheric-temperature plasma into the corona. This mechanism was verified by a simple non-LTE cloud model (Beckers 1964), where the emission of the rising cool plasma explained the blue wing enhancement, and the peak asymmetries were naturally reproduced on account of the cool upward moving material shifting the maximum opacity of the line into the blue. The peak asymmetries and blue wing enhancements were not observed in lines such as Ca ii K, Ca ii 8542 Å, and Hα, which have significantly smaller optical depths than the Mg ii lines. Additionally, using a non-LTE simulation with partial redistribution taken into consideration for the line cores, Rubio da Costa & Kleint (2017) generated synthetic blueshifted h&k lines with the desired blue and red peak asymmetries by combining two spectral profiles produced by downflows at different chromospheric heights.

We find that profiles located at the ribbon front are often the broadest profiles in the field of view, only surpassed by even broader profiles due to filament eruptions (group 10) or large downflows at the end of the flares (group 8). Large broadenings may indicate large non-thermal velocities and thus turbulence. We therefore conclude that if the profiles are resolved, the turbulence is highest at the leading edge of the flare ribbon.

Profiles from the combination of groups 11 and 12 occur in every flare, although tenuously and with extremely low triplet emissions in the succession of M-class flares on 2014 October 26. However, the slit of IRIS was positioned poorly in relation to the active region's major ribbon activity. On this day, flare 9 produced a profile assigned to group 11 with precise width and peak asymmetries, but with no triplet emission. The profile seems to be caused by upflowing material from a clearly visible erupting jet.

We conclude that the ribbon-front IRIS profiles have three possible origins: (1) They are generated by the superposition of unresolved downflows at different chromospheric heights, as demonstrated by Rubio da Costa & Kleint (2017). (2) They come about due to enhanced turbulence at the leading edge of the ribbon front, with triplet line emission due to the heating of the lower chromosphere, in line with the quiet Sun triplet modeling of Pereira et al. (2015). (3) They are generated by rising cool chromospheric-temperature material, as discussed by Tei et al. (2018). The last explanation seems unlikely, since flare 9 and the cloud model of their study generated similar profiles to our ribbon-front IRIS profiles but without the subordinate emission. This makes it clear that there should be a distinction between profiles generated from upflowing material and profiles generated from non-thermal electron bombardment. This distinction is based on the engagement of the two red wing triplet lines.