1. Introduction
Background
Tornadoes and large hail are the deadliest and costliest severe convective weather hazards in the United States, respectively (Galway 1975; Gunturi and Tippett 2017). Of all tornadoes, very few are actually deadly (Galway 1975), with those rated EF4–5 responsible for over two-thirds of these deaths (Ashley 2007). Of severe hailstorms, those that produce stones larger than 2 in. (hereafter, “large”) are particularly impactful to lives and property (Johnson and Sugden 2014; Blair et al. 2017). Clearly, though all tornadoes and hailstorms can be dangerous, higher-end events can have greater impacts. Therefore, it is prudent to understand the storm-scale evolutions preceding these events.
Both tornado intensity and maximum hail size have been found to be predictable to some extent using environmental variables. Stronger low-level shear has been associated with stronger tornadoes, both in supercells (Thompson et al. 2003; Coffer et al. 2020) and convective systems (Schenkman et al. 2011). Stronger low-level shear may also be associated with greater streamwise vorticity, which directly induces updraft rotation (Davies-Jones et al. 1990). Storm-relative helicity (SRH), a product of streamwise vorticity and storm-relative inflow, has shown appreciable skill in detecting significant tornado environments (Thompson et al. 2002, 2004; Coffer et al. 2019) though this may be due mostly to its streamwise vorticity component (Peters et al. 2023). Stronger low-level shear may also be associated with stronger low-level storm-relative winds, which can support larger storms (Peters et al. 2019, 2020, 2023). Consequently, tornadic storms are associated with stronger low-level storm-relative winds than nontornadic storms (Coniglio and Parker 2020), and potential hail size tends to increase with stronger low-level storm-relative winds (Dennis and Kumjian 2017; Kumjian et al. 2021).
Counter-intuitively, however, excessively strong storm-relative winds may make tornadogenesis more difficult, both by decreasing the efficiency of vortex tilting (Peters et al. 2023) and minimizing the residence time of incipient vortices beneath an updraft (Fischer and Dahl 2023). Similarly, although sufficiently strong storm-relative winds are necessary for hail-producing storms (Dennis and Kumjian 2017), excessively strong storm-relative winds may minimize the residence time of hail embryos within the hail growth zone, thus inhibiting hail growth (Dennis and Kumjian 2017; Kumjian et al. 2021). Upper-level storm-relative winds may also regulate tornadic activity, with excessive midlevel veering linked to downdraft placement in the forward flank inflow region of a supercell, which thwarts tornado potential in simulations (Gray and Frame 2021).
Stronger convective available potential energy (CAPE), especially at low levels, has been associated with more intense tornadoes (Hampshire et al. 2018), presumably by inducing stronger low-level updrafts and vertical vorticity stretching (Davies 2002). Conversely, excessively strong low-level CAPE may be less favorable for hail growth, as it may limit the residence time of stones within the hail growth zone (Lin and Kumjian 2022; Nixon et al. 2023). Perhaps consequently, hail has often been found along cold fronts (Schemm et al. 2016; Kunz et al. 2020), the cool side of boundaries (Magee and Davenport 2020), or “elevated” above a stable air mass (Corfidi et al. 2008).
Though a storm’s background environment is critical for severe weather forecasting, the structure of any given storm is also heavily impacted by external interactions. Interactions with airmass boundaries have frequently been noted in cases of tornadoes (Maddox et al. 1980; Wakimoto and Wilson 1989; Markowski et al. 1998a; Atkins et al. 1999; Rasmussen et al. 2000; Caruso and Davies 2005; Boustead et al. 2013; Magee and Davenport 2020) as well as large hail (Schemm et al. 2016; Kunz et al. 2020; Magee and Davenport 2020). Distinct discontinuities between air masses may serve as initiation mechanisms for convection (Schreiber 1986; Wilson and Schreiber 1986; Wakimoto and Murphey 2010), influence storm motion (Zeitler and Bunkers 2005), and provide enhanced vertical vorticity (Wakimoto and Wilson 1989; Caruso and Davies 2005) and horizontal vorticity (Markowski et al. 1998a; Atkins et al. 1999; Rasmussen et al. 2000). Boundary layer winds may also veer more rapidly on the cool side of boundaries, potentially generating more streamwise vorticity (Maddox et al. 1980; Rasmussen et al. 2000). Both tornadoes and large hail are most frequently noted on the immediate cool side of boundaries (Markowski et al. 1998a; Atkins et al. 1999; Rasmussen et al. 2000; Schemm et al. 2016; Kunz et al. 2020; Magee and Davenport 2020) though tornadoes are often produced closer to the boundary (Markowski et al. 1998a; Magee and Davenport 2020).
Traditionally, the majority of both significant tornadoes and large hail have been thought to be produced by discrete supercells (Smith et al. 2012; Blair et al. 2017). That said, cell mergers have been found to precede tornadogenesis quite frequently (Lee et al. 2006; Rogers and Weiss 2008; Rogers 2012; Broyles et al. 2022; Flournoy et al. 2022; Fischer and Dahl 2023; Lyza and Flournoy 2023) and have also been observed in cases of very large hail (Piasecki et al. 2023). Some mergers precede mesocyclone intensification (Flournoy et al. 2022; Lyza and Flournoy 2023) and may strengthen a storm by directly contributing to its updraft (Hastings and Richardson 2016) or by acting against the low-level storm-relative wind to enhance convergence along its gust front (Jewett et al. 2006; Fischer and Dahl 2023). Convective systems are also known to frequently produce tornadoes (Trapp et al. 2005; Smith et al. 2012; Ashley et al. 2019), albeit primarily weak (EF0–EF1) ones. In these systems, stronger environmental low-level shear may encourage stronger and more upright updrafts (Schenkman et al. 2011), and enhanced convergence along the gust front may similarly assist tornadogenesis (Rotunno and Klemp 1985; Atkins and St. Laurent 2009; Schenkman et al. 2012). Embedded supercells within convective systems may carry a heightened threat for tornadoes (Barker 2006; French and Parker 2012; Smith et al. 2012), but their ability to produce large hail is less understood.
Warning criteria for hail (Donavon and Jungbluth 2007; Blair et al. 2011) and tornadoes (Smith et al. 2015; Gibbs and Bowers 2019) are based primarily on radar observations of the current state of the storm in question, as well as some environmental parameters. There is ample evidence (e.g., Lee et al. 2006; Rogers 2012; Hastings and Richardson 2016; Flournoy et al. 2022; Fischer and Dahl 2023; Lyza and Flournoy 2023) that a better understanding of storm interactions may provide insight into a storm’s potential future state, which could be helpful for warning decisions. Particularly, both longer tornado lead times (Brooks and Correia 2018) and forecasts of maximum expected hail size (Blair et al. 2017) may be possible with a better understanding of storm interactions. Currently, external interactions present a confounding variable to our ability to anticipate hazard potential, especially because they can occur independently of the background environmental conditions, and thus are not predictable using parameter-based forecasting alone.
To better understand the importance of cell mergers, boundary interactions, and convective systems in hazard production, we use WSR-88D data to investigate the following:
-
What types of external features are most commonly observed to precede significant tornadoes and large hail?
-
Do certain features become more or less common depending on the environment?
-
How might these features be important for hazard production?
To answer these questions, we identify a variety of external features that may have interacted with the storm in question and present conceptual models for their positions preceding tornadoes and large hail. This design differs from that of most previous studies, which primarily focused on one hazard [with the exception of Magee and Davenport (2020)] or one type of interaction in isolation [e.g., only cell mergers or only boundaries, with the exception of Fischer and Dahl (2023)]. With this detailed examination of potential storm interactions, we hope to provide additional insight into this complex forecasting problem.
2. Data and methods
a. Case selection
The cases used in this study were collected from NOAA NCEI Storm Data reports (NOAA/NCEI 2008) from 2008 to 2021. All large hail events were considered that were at least 3 in. (hereafter “very large”). All tornadoes were considered that were rated at least EF3. To build a more comparable sample size between tornado and large hail events, tornado cases were added from 2022 to 2023. In less populated areas, both maximum hail size (Dobur 2005; Blair and Leighton 2012; Groenemeijer et al. 2017; Allen et al. 2020) and maximum tornado rating (Doswell and Burgess 1988; Wurman et al. 2021) are less likely to be recorded. To avoid possibly underrepresenting significant tornado frequency in the Great Plains due to a lack of damage indicators (Doswell and Burgess 1988; Wurman et al. 2021), we chose to include EF2 tornadoes west of 98° longitude, where population density becomes sparse (there were 65 such cases). This threshold was determined arbitrarily, informed by county-level U.S. Census population data (U.S. Census 2020), and may not reflect the actual availability of damage indicators for each case (or a tornado’s potential to be stronger than EF2). We hereafter refer to these EF2+ tornadoes as “significant” (Hales 1988). Both right-moving (RM) and left-moving (LM) supercells were considered for this study; 27 cases of very large hail were produced by left-moving supercells. That said, a supercell storm was not a necessary requirement for this dataset. While all cases of tornadoes examined were by nature produced by storms with detectable mesocyclones at the time of the event, 17 cases of very large hail occurred with storms lacking an identifiable mesocyclone.
The coordinates of the recorded tornadogenesis point or hail report must also have occurred within 90 km of a WSR-88D (Crum and Alberty 1993). This distance was chosen arbitrarily to optimize the ability to examine any low-level features in detail while also maximizing sample size. In rare instances, cases that were within these radii were still omitted if the cells or features in question were too small or distant to distinguish from one another (e.g., although beam resolution was high enough to distinguish between multiple large, widely spaced supercells at 90 km, it may not have been enough to distinguish between multiple closely spaced, smaller cells at 90 km).
We assume all events were reported accurately in location and time (though this is not always the case, e.g., Blair et al. 2017). An event was considered suspect if (for tornadoes) it did not occur within 15 min of the first scan that featured a tornadic vortex signature (TVS, Brown et al. 1978; Mitchell et al. 1998) or (for very large hail) it did not coincide with a region of at least 55 dBZ [as in Czernecki et al. (2019)] within 5 km of the report location. For any suspect events, the report time stamp was adjusted if possible, else the case was omitted. To avoid overrepresenting certain environments, duplicate reports (defined as occurring within 2 h of another report nearest to the same WSR-88D) were flagged, and only the report with the highest magnitude was considered as a case. In the case of duplicate magnitudes, if the report was within 70 km of the WSR-88D, only the earliest time was considered. Outside 70 km of the WSR-88D, only the closest report was considered. In total, 246 3 in.+ hail cases and 230 EF2+ tornado cases (476 total cases) are examined in this study.
For each report of very large hail and significant tornado, when possible, a null case was also obtained at the time of that report from the nearest rotating storm that failed to produce that hazard within 30 min of the report. For instance, if a very large hail report occurred at 2200 UTC from one storm, the closest storm that did not produce any severe hail reports between 2130 and 2200 UTC was considered to be a null case. Note that a null case storm for hail may still have produced subsevere hail, and we assume that underreporting of hailstones does not affect case collection. A null case storm must have been within 90 km of the same WSR-88D that the hazardous storm was observed from (thus close enough to the hazardous storm that both storms could be assumed to exist in similar environments). The null case storm must have had a detectable mesocyclone at any level for at least 20 min and have the same direction of rotation as the storm that did produce the hazard. In total, 107 null cases of tornadoes and 84 null cases of severe hail are examined in this study.
b. Environmental data
Environmental data were obtained for each case from the fifth major global reanalysis produced by ECMWF (ERA5; ECMWF Climate Data Store 2017; Hersbach et al. 2020), which is known to reliably represent severe convective environments in the United States (Coffer et al. 2020; Taszarek et al. 2021). Vertical profiles were taken from 137 hybrid-sigma model levels of ERA5 data at the nearest grid point to the latitude and longitude of each report and at the nearest time to the report. All profiles were interpolated to a common 250-m vertical resolution before compositing. The various kinematic and thermodynamic variables used in this study were derived from these data profiles using MetPy (May et al. 2022). Thermodynamic parameters include CAPE, CAPE in the lowest 3 km (3CAPE), and convective inhibition (CIN) (J kg−1), for the most unstable (MU), 100-mb mixed-layer (ML; 1 mb = 1 hPa), and surface-based (SB) parcels, depending on application. Lifting condensation level (LCL) was calculated using the ML parcel. Kinematic parameters include 0–500-m bulk wind difference (BWD) and mean 0–500-m storm-relative inflow (m s−1). This storm-relative inflow was calculated using the observed storm motion in tornadic cases (by tracking the difference in position of the lowest-tilt mesocyclone from T − 20 to T − 0 min before hazard onset) and the Bunkers right-moving storm motion estimate (Bunkers et al. 2000) for all other cases. Like other reanalyses, known biases in ERA5 data are confined mostly to the boundary layer (Allen et al. 2014; Gensini et al. 2014; Taszarek et al. 2018; King and Kennedy 2019). Because cases often occur near surface boundaries in this dataset, the ERA5 data obtained for each case may not always accurately depict its local environment, especially given its 0.25° spatial resolution. While this may prevent definitive conclusions from being drawn about what environment a storm is truly experiencing in these cases (Allen and Karoly 2014; Coniglio and Jewell 2022), we examine these cases in the mean sense and discuss uncertainty where applicable.
c. Examination of WSR-88D data
Level II data from the closest WSR-88D to each tornadogenesis point or hail report were downloaded and analyzed using the Gibson Ridge Level II (GRLevel2) Analyst software (Gibson Ridge 2005). Though past studies often examined data up to 15 min preceding the event (e.g., Lee et al. 2006; Rogers and Weiss 2008; Fischer and Dahl 2023), this study uses a 30-min study window to 1) account for cases where an earlier interaction had significant impacts on storm characteristics (e.g., Hastings and Richardson 2016) and 2) better understand how storm interactions may change closer to an event. Cases were occasionally examined prior to this period in order to gain more context (for instance, to determine whether a surface boundary was produced by the storm in question or had already existed prior to its initiation). Wherever referenced herein, a storm’s precipitation footprint was defined as the area of reflectivity of at least 30 dBZ (Smith 2010; Smith et al. 2012) and its core the region bound by the highest 15 dBZ reflectivity within this footprint. A storm’s centroid was defined as the center of its core. An example of these attributes is illustrated in Fig. 1: storm identification.
An example showing (left) the identification of storm features and (right) the positions of other cells. Shown is base reflectivity data (dBZ) from the Aberdeen, SD (KABR), WSR-88D on 22 May 2010, plotted using GRLevel2. The storm’s mesocyclone is marked with a black triangle, and its 30-dBZ precipitation footprint is circled in solid white. In the left panel, the storm’s precipitation axis, highest 15-dBZ core, and centroid are labeled. In the right panel, the storm’s precipitation axis is juxtaposed with its mesocyclone (this is its orientation), and its forward, rear, right, and left flanks are labeled. A neighboring cell (its 30-dBZ precipitation footprint and centroid) is circled.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
The storm-relative positions of external features were of particular priority to this study. Ground-relative information (i.e., southeast or northwest) was insufficient for this study since storms assumed a wide variety of orientations (e.g., a storm’s rear flank was not always to the southwest of its precipitation core). The motions of storms and their interactions were also not enough since 1) a storm’s structure depends more directly on storm-relative winds than its motion and 2) a storm may interact with other features that have no motion relative to the storm. Consequently, we consider the storm-relative direction and motion of these features to be important. To assess these, we consider the orientation of a storm and its flanks. A storm’s orientation was defined as the direction toward which its precipitation axis pointed (e.g., a precipitation footprint that stretched from the southwest to the northeast would have a northeast orientation). This precipitation axis was recorded as the mean angle of the line connecting the ends of the major axis of the storm’s precipitation footprint (excluding the hook or pendant echo) in the 15-min preceding hazard production (shown in Fig. 1: storm identification). Accordingly, a storm’s forward flank can be defined as the region downwind of a storm with respect to its orientation (between 315° and 45° clockwise from its orientation, as in Fig. 1: cell positions), a storm’s rear flank upwind (135°–225°), its right flank to the right (45°–135°), and its left flank to the left (225°–315°). This methodology is similar to that employed by Lyza and Flournoy (2023) but with precise and predetermined bounds such that could be applied for a variety of different supercell shapes in different environments. The right and left flanks were mirrored for left-moving supercells. All detectable external features were noted using the nomenclature described in the following subsection.
d. Identification of external features
All features were identified manually by the first author, according to the categorical system described in this section. Multiple passes were made through the dataset to improve confidence in detection and consistency of identification.
Features were distinguished by type:
-
Boundary: fine-line airmass boundary
-
System: line of conjoined cells
-
Supercell: cell with mesocyclone
-
Cell: cell without mesocyclone
We partition the analysis into two main questions. First, where do tornadic storms and hailstorms tend to be positioned relative to boundaries, systems, and other supercells? Second, where do cell mergers and other nearby cells tend to be positioned relative to storms before they produce tornadoes or large hail? We identify potential interactions as follows. We strongly recommend the reader to peruse the online supplemental material for examples of all features that follow and to assist reproducibility.
1) Boundaries
A boundary was recorded when a storm was within 50 km of a fine-line signature on the lowest tilt scan (Markowski et al. 1998a). These boundary layer convergence lines are regularly noted with cold fronts, drylines, outflow boundaries, and differential heating zones (Schreiber 1986; Wilson and Schreiber 1986; Wakimoto and Wilson 1989; Wakimoto and Murphey 2010). These boundaries must have been noted to exist prior to the formation of the storm in question (i.e., not the storm’s own outflow boundary). A boundary must have been detectable on radar to be included in this study. This is imperative in order to ensure that its position relative to a storm is known. We acknowledge that, although this study uses a 90-km maximum distance from a WSR-88D in order to detect shallow features such as boundaries, some boundaries may still be too shallow to detect. Furthermore, some boundaries, like warm fronts, may be too diffuse to detect (Markowski et al. 1998b; Garner 2013; Magee and Davenport 2020). The availability of scatterers may also affect a boundary’s detectability (Wilson and Schreiber 1986). The exclusion of boundaries that are more likely to be weak and diffuse is considered sufficient for the purpose of this study.
A storm’s position relative to the boundary was of primary importance since this presumably had implications as to whether a storm’s inflow air was from the more-unstable or less unstable side of the boundary. This assessment depended on analysis using archived Storm Prediction Center mesoanalysis data (NWS Storm Prediction Center 2005), where the air mass with smaller ML3CAPE was denoted as “stable,” and vice versa (note that although we use this naming convention, the side of the boundary with smaller ML3CAPE may not necessarily be stable but relatively less unstable). If ML3CAPE was less than 50 J kg−1 on both sides of the boundary, mixed-layer CAPE (MLCAPE) was used. If MLCAPE was insufficient to resolve the boundary (e.g., a boundary was detected on radar but could not be detected using this field, which was rare), radar cues defining the boundary were evaluated. For instance, a storm was considered to be on the unstable (stable) side of a boundary if that air mass featured stronger (weaker) clear-air echo returns and more widely spaced (tightly spaced) horizontal convective rolls, an indication of boundary layer turbulence and instability (Christian and Wakimoto 1989; Weckwerth et al. 1997). Whether or not a storm was collocated with a boundary was also noted since this may have provided it with enhanced near-surface baroclinic vorticity (Markowski et al. 1998a; Atkins et al. 1999; Rasmussen et al. 2000). Whether or not a storm collided with a boundary was also noted.
A storm’s position near a boundary was defined as follows:
-
Unstable side: The storm’s mesocyclone was within unstable air mass,1 or the storm was oriented into unstable air mass.
-
Stable side: The storm’s mesocyclone was within stable air mass, or the storm was oriented into stable air mass (“undercut,” as in Markowski and Richardson 2009).
-
Anchored: The storm’s mesocyclone was within 10 km of the boundary for at least 20 min [thus, its motion was presumed to be influenced by the boundary, as in Zeitler and Bunkers (2005)].
-
Collision: The boundary collided with the storm’s mesocyclone.
2) Systems
A system was recorded when the storm was part of a continuous line of cells of at least 30-dBZ reflectivity, composed of at least some nonsupercellular cores, that extended at least 50 km long. Though most convective systems that fit into this category were traditional quasi-linear convective systems (Trapp et al. 2005), the shorter length criteria allowed for smaller systems to be incorporated, such as supercells that developed extensive trailing convection along their gust fronts. Though these systems blurred the lines between the scales of mesoscale convective systems and supercells, they were included in this category because they possessed a more substantial cold pool that initiated convection.
Similarly to boundaries, a storm’s position within a convective system was of primary importance since this presumably had implications on whether a storm’s inflow air was from the more unstable warm sector or less unstable outflow from other storms. Whether or not a storm collided with a convective system was also noted.
A storm’s position within a system was defined as follows:
-
Head: Line of cells extended into the storm’s rear or right flank, but not its forward or left flank [e.g., the storm was the northernmost cell in a line, occasionally referred to as the “bookend vortex” (Weisman 1993)]
-
Tail: Line of cells extended into storm’s forward or left flank, but not its rear or right flank [e.g., storm was the southernmost cell in a line, or “tail-end Charlie” (Branick 1996)]
-
Embedded: Lines of reflectivity extended into both the storm’s forward and rear flanks or right and left flanks
-
Unstable side: The storm’s mesocyclone was ahead of the system’s outflow boundary, or the storm was oriented into the air mass ahead of the outflow boundary.
-
Stable side: The storm’s mesocyclone was behind the system’s outflow boundary, or the storm was oriented into stable air mass behind the outflow boundary.
-
-
Collision: The system collided with the storm’s mesocyclone.
3) Supercells
A supercell was recorded when another supercellular mesocyclone was within 50 km of the mesocyclone of the storm in question. Past studies such as Klees et al. (2016) and Wilson et al. (2023) have noted that two supercells in close proximity to one another may affect each other’s local environments and their abilities to produce hazards. Although supercells can merge or collide, we also considered supercells that simply coexisted within 50 km of one another for at least 10 min. For these cases, we considered only supercells of the same direction of rotation (since these tended to have slow or negligible motions relative to each other). If there were multiple supercells, the closest supercell was considered.
A supercell’s position relative to another was considered important since this presumably had implications on how each storm was modifying the other’s inflow air. Whether or not a supercell impacted another supercell of the same direction of rotation (a “merger”) or opposite direction of rotation (a “collision”) was also noted. The dynamical differences between the impacts of “mergers” and “collisions” are assumed to be nontrivial; while interactions between RM and RM supercells were often slow and typically resulted in the combination of both precipitation footprints (and even mesocyclones) into one coherent supercell, interactions between RM and LM supercells were often sudden and typically did not result in such a coherent combination. In that way, they resembled collisions with convective systems or boundaries.
A storm’s position near another supercell was defined as follows:
-
Head: Another supercell’s mesocyclone (that had the same direction of rotation) was in the rear flank of the storm.
-
Tail: Another supercell’s mesocyclone (that had the same direction of rotation) was in the forward flank of the storm [e.g., the storm was the southernmost cell in a pair, or “tail-end Charlie” (Branick 1996)].
-
Rightmost (leftmost): Another supercell’s mesocyclone (that had the same direction of rotation) was in the left (right) flank of the storm.
-
Merger: The storm’s mesocyclone and/or precipitation footprint merged with another supercell of the same direction of rotation such that they became indistinguishable from one another.
-
Collision: The storm was impacted by the precipitation footprint or gust front of another supercell of the opposite direction of rotation.
Examples of these scenarios are included in Figs. S14–S20.
Occasionally, in interactions between two supercells, it was not obvious which storm was the storm in question. Whenever possible, the storm in question was the storm that produced the tornado within its mesocyclone or pendant echo or the very large hail within its precipitation core. However, when two supercells fully merged or collided such that their mesocyclones and precipitation cores became indistinguishable from one another, this could not be determined. In these cases, the storm in question was the storm that had the closest motion vector to the initial motion of the tornado (or the motion of the reflectivity core that produced the large hail). If this could not be ascertained (which was very rare), the storm in question was the storm with the highest-reflectivity core in the 15 min preceding the hazard. Examples of this scenario are also illustrated in Figs. S16–S20.
Additionally, when a storm generated more than one mesocyclone, it was not always obvious whether there were one or two supercells being analyzed. Mesocyclones commonly “hand off” (e.g., Edwards 2014) in cyclic supercells during the occlusion process (Adlerman et al. 1999; Dowell and Bluestein 2002), and calling these two supercells is not conventional. However, in rare instances, a discarded mesocyclone persisted and matured, and its attendant precipitation footprint became distinguishable from that of its parent supercell. In other instances, a storm generated a new mesocyclone separate from the original such that the two were never conjoined. This did not fit the conceptual model of occlusion, but rather discrete propagation [as in Zeitler and Bunkers (2005)]. In any case, regardless of the presumed mechanism, if two mesocyclones with distinguishable precipitation footprints were noted, this was treated as two supercells. Examples of these scenarios are included in Figs. S21 and S22.
4) Neighboring cells
All neighboring cells were recorded when any part of their precipitation footprint was detected within 25 km of the mesocyclone2 of the storm in question. Although studies such as Fischer and Dahl (2023) used a smaller radius (in their case, 5 km) to search for nearby cells, we used a longer radius to account for scenarios where the cold pools of these cells [or the pressure perturbations from their outflow (Fischer and Dahl 2023)] propagated away from their actual precipitation footprints because these cold pools could not easily be detected by WSR-88D. Since the impacts of cell interactions have been found to be highly variable and depend on their maturities and distances from the storm (Hastings and Richardson 2016), we incorporated both of these considerations into the search process. Cells were only considered if their precipitation core achieved 30-dBZ reflectivity for at least 10 min. Furthermore, in an effort to consider only the most potentially impactful cells, if there were multiple cells all in the same direction from the storm, we devised a method to consider only the closest of these storms. If a farther cell’s precipitation centroid was eclipsed by the precipitation footprint of a closer cell from the line of sight of the storm’s mesocyclone, only the closer cell was considered. This methodology in a particularly complex example is shown in Figs. S1 and S2.
The exact storm-relative positions of the centroids of all cells (their directions and distances) relative to the storm’s mesocyclone were recorded whether they be nearby cells, nearby supercells, or cores within a conjoined convective system. Though most of these cells originated from outside of the storm in question, many appeared to be initiated by the storm itself, including cells along the rear-flank gust front, splitting cells, and, rarely, occluded updrafts. Though distinct cell mergers are most commonly cited in literature, we chose to also consider all nearby cells that were not part of the precipitation footprint of the storm in question, even if they merged only partially, or did not merge, as long as they remained near the storm for at least 10 min. Cells were considered connected if their 30-dBZ precipitation footprint intersected that of the storm in question. Analyses were performed 0, 10, 20, and 30 min prior to hazard production. In the occasional instance where the storm in question did not yet have a mesocyclone or identifiable pendant echo at longer lead times (e.g., 30 min prior to hazard production), analysis ceased, and no neighboring cells were recorded at these times.
Cells were defined as follows:
-
Split: A previously connected cell that split from the storm such that its precipitation footprint became distinguishable from that of the storm (as in Klemp and Wilhelmson 1978)
-
Merger: A previously unconnected cell that fully merged with the storm such that its precipitation footprint became indistinguishable from that of the storm (Westcott 1984; Lee et al. 2006)
-
Nudger: An unconnected cell that did not merge with the storm
-
Feeder: A connected cell that remained partially merged with the storm [similar to Zeitler and Bunkers (2005)]
Examples of these scenarios are included in Figs. S23–S28.
On rare occasions, none of the features in this section were recorded near a storm prior to hazard production. Examples of noninteracting storms are included in Figs. S29 and S30.
We acknowledge that storm-scale interactions are an extremely complex topic of evolving understanding. A plethora of other heterogeneities may also impact a storm, such as subtle density currents, anvil shading, and terrain roughness. That said, this study will focus on the features that are discernible using WSR-88D data, which is readily accessible in operations. We also acknowledge that results may be sensitive to how features are defined and interpreted by other people. For instance, a storm’s orientation and the position of its flanks can be defined in multiple ways. Herein, we define it based on the direction of the major axis of the storm’s precipitation footprint, which may be interpreted differently by different people. Consequently, the positions of nearby cells relative to a storm, and which flanks they occupy, may not be analyzed the same way across multiple people. While analysis by a single person can be optimal for consistency across the data evaluated, reproducibility may be improved by analysis by multiple people, as long as each person analyzes every case in the dataset. Otherwise, machine learning may be useful in future work.
3. Results
a. Types of features preceding each hazard
First, we examine the types of features noted in the 15 min prior to tornadogenesis and hail production, in order to gain context into their contributions to the broader climatology of hazard production. Storms producing significant tornadoes and very large hail were almost always accompanied by external features in the 15-min preceding hazard production, with fewer than 10% lacking a nearby boundary, convective system, supercell, or neighboring cell (Fig. 2).
Bar charts showing the percentage of cases in which each feature was observed in the 15 min preceding (top left) null tornado events, (top right) significant tornadoes, (bottom left) null hail events, and (bottom right) significant hail events. Bars are colored by the type of interaction. Percentages are relative to the total number of cases in each sample. Note that because neighboring cells were so numerous in each sample (and occurred alongside boundaries, systems, and supercells), only the percentage of cases that had only neighboring cells is shown. Cases in which no features were noted are labeled “None.”
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Significant tornadoes were most commonly associated with convective systems, neighboring supercells, and other cells and supercells (Fig. 2). Convective systems accounted for over 30% of all cases. Though they have traditionally held a reputation for producing primarily weak tornadoes (Trapp et al. 2005; Smith et al. 2012), recent work has noted that they may be responsible for more significant tornadoes than previously thought (Ashley et al. 2019). Most cases in this sample were produced by supercells embedded within convective systems. Neighboring supercells were found in similarly large portions of the sample, as was also explored in Beveridge et al. (2019) and Fischer and Dahl (2023). Although boundary interactions are traditionally emphasized as a precursor to tornadoes (Wakimoto and Wilson 1989; Markowski et al. 1998a; Rasmussen et al. 2000), they appeared to be relatively uncommon in this sample of significant tornadoes and not a distinguishing factor between tornadic and nontornadic cases. We remind the reader that the strict criteria for boundary identification used herein may exclude warm fronts, which are often diffuse (Gidel 1978; Hoskins and Heckley 1981; Wakimoto and Bosart 2001; Banacos and Schultz 2005). None of the features preceding significant tornadoes appeared to vary notably in frequency from the null sample (Fig. 2), although neighboring cells and supercells were generally more numerous preceding significant tornado cases. This is to say that in general, the mere presence of convective systems, boundaries, and even neighboring cells was not a distinguishing factor between storms that produced tornadoes and those that did not.
Significant hailstorms were most commonly associated with boundaries though convective systems, supercells, and other cells were also relatively common (Fig. 2). These boundaries were found in 40% of all cases and were almost twice as common preceding significant hailstorms than in null cases. This is to say that boundaries may be a distinguishing feature of significant hailstorms. Similarly to tornadoes, the presence of convective systems, supercells, and other cells did little to distinguish between hazardous and nonhazardous storms.
A map of where the significant tornado and hail events occurred can be found in Fig. 3. As expected given their climatological frequency (Concannon et al. 2000), the significant tornado cases examined spanned a broad area over the central United States. While tornadoes associated with convective systems were relatively common throughout the United States, tornadoes near boundaries were found almost exclusively in the Great Plains. While field studies based in the Great Plains recognized the importance of surface boundaries in tornado climatology (Markowski et al. 1998a; Rasmussen et al. 2000), there may be evidence here that similar boundaries are less common outside of the plains. That said, the absence of boundary cases east of the Mississippi River may be due in part to a greater prevalence of shallower or more diffuse boundaries (such that were undetectable by WSR-88D) or simply because this region often lacks discernible surface boundaries to begin with in cases of severe weather (Chasteen and Koch 2022). Indeed, tornado events in the eastern United States are frequently driven by more subtle, elevated forcing mechanisms such as cold fronts aloft (Browning and Monk 1982; Hobbs et al. 1990, 1996; Chasteen and Koch 2022). When surface boundaries are present, they tend to be cold frontal in nature and produce large convective systems (McAvoy et al. 2000; Davis and Parker 2012; Sherburn and Parker 2014) though drylines are occasionally observed (Duell and Van Den Broeke 2016). That these convective systems contribute to an especially large portion of events in the eastern United States reflects previous studies (e.g., Trapp et al. 2005; Smith et al. 2012; Ashley et al. 2019).
Map showing (top) the genesis locations of each significant tornado and (bottom) the locations of each very large hail report examined, colored by the type of feature that was present, with a focus on boundaries and convective systems.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Hail events, on the other hand, were overwhelmingly located in the southern and central Great Plains (Fig. 3), also consistent with climatology (Allen and Tippett 2015; Allen et al. 2017; Taszarek et al. 2020; Murillo et al. 2021). In this sample, both boundaries and convective systems were present in very large hail events throughout the United States, but the frequency of boundary interactions increased substantially in the Great Plains (similar to the tornado sample). Relative to the more subtle forcing mechanisms often found in the eastern United States (Browning and Monk 1982; Hobbs et al. 1990, 1996; Chasteen and Koch 2022), these surface boundaries may play a nontrivial role in the prevalence of very large hail in the Great Plains compared to other regions (Allen and Tippett 2015; Allen et al. 2017).
b. Storm-relative positions of boundaries, systems, and supercells
Although the presence of a boundary, convective system, or nearby supercell alone did little to distinguish hazardous storms, the position of a storm relative to these features may be considerably important in the 15 min preceding hazard production. Where boundaries were noted, significantly tornadic storms were overwhelmingly found on the unstable side, while nontornadic storms were typically instead on the stable side (Fig. 4). Although this seemingly contradicts observations that tornadoes are most common on the “cool” side of boundaries (e.g., Markowski et al. 1998a; Magee and Davenport 2020), we draw our conclusions herein from derived CAPE, not temperature. Temperature may not always reflect CAPE; indeed, in the instance of most drylines and some outflow boundaries, the relatively cool side may have more CAPE because it has greater boundary layer moisture (Hanft and Houston 2018). This conclusion may also depend on the potential exclusion of warm fronts, as suggested earlier. In stark contrast to tornadic storms, hail-producing storms were overwhelmingly found on the stable side of surface boundaries, reflecting previous studies (e.g., Schemm et al. 2016; Kunz et al. 2020; Magee and Davenport 2020). In other words, these storms were likely “elevated” (Corfidi et al. 2008). This is important because it suggests that hailstorms are frequently accompanied by locally more stable boundary layers that may not be evident in ERA5 reanalysis (or operational NWP). Though null cases were also primarily produced by storms on the stable side of a boundary, we note that this sample was dominated by cases where all nearby cells were also on the stable side, so a null case was selected from the stable side by default.
(left) Pie chart showing the relative frequencies of storms’ positions near boundaries for (top left) null tornado events, (top right) significant tornado events, (bottom left) null hail events, and (bottom right) very large hail events, in the 15 min preceding each event. Positions that assume stable air in the forward flank of a supercell are colored in blue, while positions that assume stable air in the rear flank of a supercell are colored in red. Percentages are relative to the total number of cases associated with boundaries. (right) A conceptual model for boundary interactions preceding (top) significant tornado events and (bottom) significant hail events. A hypothetical average supercell (its 30-dBZ precipitation footprint) is drawn. Note that tornadic supercells tend to be oriented into an unstable air mass, such as ahead of a dryline or modified outflow boundary (brown scalloped line), while hail cases tend to be oriented into a stable air mass, such as behind a cold front or new outflow boundary (blue line with triangles).
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Within convective systems, tornadic storms were most common at the head of the line or otherwise embedded (such that they were oriented into the unstable air ahead of the gust front), while nontornadic storms were more common at the tail end or oriented into the stable side behind the gust front (Fig. 5). Again in contrast, hailstorms were found most often at the tail end of a line or otherwise oriented into the stable side, while null cases were more common at the head or oriented into the unstable side. This mirrors the relationship found in Fig. 4, where most hailstorms were likely elevated above the system’s cold pool, while tornadic cases had unstable air in their forward flanks.
As in Fig. 4, but for the relative frequencies of storms’ positions within convective systems. Note that tornadic supercells tend to be at the head of a line (or embedded within it), while hail cases tend to be at the tail of a line.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
The position of supercells in a pairing of supercells was a similarly distinguishing factor for tornado potential, where tornadic supercells were predominantly ahead or to the right of another supercell, while nontornadic supercells were at the tail or left of another supercell (Fig. 6). This is counter to the traditional (and anecdotal) thought that the southernmost supercell [or “tail-end Charlie” (Branick 1996)] carries the highest tornado potential. Indeed, Beveridge et al. (2019) found that the north-relative position within a line of storms has no statistically significant impact on tornado potential. However, we do find considerable differences in the positions of tornadic versus nontornadic storms in this sample, perhaps because we examine orientation as storm-relative rather than north-relative. Hailstorms, conversely, were predominantly at the tail end of another supercell, while null cases were at the head of another supercell. These findings again echo those examined in Figs. 4 and 5, where tornadic supercells had access to unstable air in their forward flanks, while hailstorms had locally more stable or rain-cooled air in their forward flanks. Neither significant hail nor tornadoes were common with supercells to the left of other supercells.
As in Fig. 4, but for the relative frequencies of storms’ positions near other supercells. Note that tornadic supercells tend to be at the head or right of another supercell, hail cases tend to be at the tail of another supercell, and null cases are often at the left of another supercell.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
c. Storm-relative positions of neighboring cells
Although neighboring cells frequently accompany both tornadic storms and hailstorms, the positions of these cells relative to a storm may be of particular importance in anticipating hazard production (or lack thereof). Tornadic supercells were frequently accompanied by other cells in their left and rear flanks (Fig. 7). These cells were most commonly found near the storm’s hook echo, in a similar position as was found in Broyles et al. (2022, their Figs. 23 and 26). Their most favored position was generally opposite the direction of the mean 0–500-m storm-relative wind (with a density peak at 195° from this vector). The potential implications of this are discussed in Fischer and Dahl (2023) and later in the summary and discussion section. Tornadic storms had more cells in their rear flanks and left flanks than nontornadic storms, especially at T − 30, which suggests that these cells may be an indicator of tornado potential at longer lead times. Although nontornadic storms frequently had rear-flank cells at T − 10, they were less concentrated in the area directly opposing the storm-relative wind. In tornadic storms, the frequency of cells in the left flank decreased by T − 10 (this was not true for nontornadic storms). In other words, tornadic storms featured more cells in their left flanks in the 30 min prior to tornadogenesis, but most of these cells either merged or dissipated before tornadogenesis.
Neighboring cell positions relative to the parent storm’s mesocyclone at (top) T − 30 and (bottom) T − 10 min before (left) null cases and (right) significant tornado cases. The origin is marked by a black dot. A hypothetical supercell’s 30-dBZ reflectivity contour is contoured in black. All neighboring cells are plotted as blue dots. The storm-relative positions of cells for each case are rotated such that their corresponding storm’s precipitation axis lies on the positive y axis. The normalized densities of these cells (unitless) are plotted in red. Densities were calculated using a 2D histogram of cell positions with 30 bins and then normalized by sample size to be comparable between null cases and hazard cases. Higher normalized densities in a given position imply that neighboring cells more frequently accompanied the storm there. These densities were smoothed using a 1.25-sigma Gaussian filter and displayed above a threshold value of 1.0 to de-emphasize infrequent cell positions. Range rings denote the distance from the mesocyclone, and grid lines delineate the bounds of the storm’s flanks (forward, left, etc.). The ERA5 mean 0–500-m storm-relative wind (using the Bunkers right vector for null cases and radar-estimated vector for all other cases) is plotted as a wind barb (knots).
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Another distinctive difference between tornadic and nontornadic storms was the presence of cells in the forward flank. While null cases featured forward-flank cells quite commonly, tornadic storms did not. This was also evident in the results presented in Figs. 5 and 6. The positions of these forward-flank cells in nontornadic cases resembled the distribution of downdrafts that were unfavorable for tornadogenesis in Gray and Frame (2021). In that study, downdrafts more directly upstream with respect to the low-level storm-relative inflow were more likely to contaminate the storm’s low-level mesocyclone region with more stable air. We believe that the storm-relative inflow direction is similarly important in these cases of external downdrafts, whereupon cells more directly upstream of a supercell (with respect to its storm-relative inflow) are less favorable for significant tornadoes. Traditionally, the southernmost supercell (or “tail-end Charlie”; Branick 1996) is thought to pose the greatest tornado potential because it has the most unimpeded access to buoyant inflow air to its south (Kuster et al. 2015). However, we caution that the direction of storm-relative inflow varies in different shear profiles, and we find that more generally, storms with greater tornado potential have fewer cells upstream with respect to this inflow.
Hailstorms, in contrast to tornadic storms, were accompanied by fewer cells on average and featured more cells in their forward and left flanks than their rear flanks (Fig. 8). At T − 10 min, forward-flank cells were more common preceding very large hail cases than null cases, which were also seen in Figs. 5 and 6. Although hailstorms frequently featured cells in their left flanks, few cells were found directly rearward of the storms. Compared to the tornadic cases, the presence of left-flank cells was relatively sparse at T − 30 and appeared not to distinguish between hail-producing and non-hail-producing storms (at any lead time). Few cells occupied the region directly opposite the mean storm-relative inflow in very large hail cases. We note, however, that because a large proportion of hailstorms were north of outflow boundaries and sharp cold fronts that were likely not resolved by ERA5 (as noted in Fig. 11), this storm-relative inflow direction may not be representative of what the storm was actually experiencing, so few conclusions are made about this going forward.
As in Fig. 7, but for hailstorms. The normalized densities are comparable to those in Fig. 7, owing to a random sampling of hail cases to ensure an equal sample size to tornado cases. The ERA5 mean 0–500-m storm-relative wind (using the Bunkers right vector for all cases) is plotted as a wind barb (knots).
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
In general, there were relatively few neighboring cells at T − 30 and little difference in their densities between hail cases and null cases at this time. This suggests that neighboring cells may not be a good indicator of hail potential at longer lead times. Although neighboring cells may have less of an obvious relationship to large hail potential in general than they had to tornado potential, we remind the reader that more than 75% of very large hail cases featured boundaries, convective systems, or nearby supercells, all of which could distinguish between hail-producing and non-hail-producing storms (Figs. 4–6).
d. Mergers and collisions preceding each hazard
Now that we have a better understanding of where external features tend to be positioned relative to a storm, we examine how frequently these features merge, collide, or split with the storm in question before hazard production. We remind the reader that mergers and collisions are assumed to be different; while mergers were coherent combinations of storms (such as two RM supercells), collisions involved sudden impacts between dissimilar features (such as an RM supercell with an LM supercell, boundary, or convective system, as described in the data and methods section—Identification of external features). In the 15 min preceding hazard production, significant tornadoes were associated with more of these events than hailstorms, with almost 50% of cases experiencing an event (Fig. 9). While merging cells and supercells were more common preceding significant tornadoes, splitting cells were less common. Significant hail cases had notably fewer instances of cell and supercell mergers than tornadic cases, at least in the 15 min preceding hazard production [but we note that very large hail has still been noted to follow cell mergers, as in Piasecki et al. (2023)]. Hail cases also featured the greatest frequency of collisions from boundaries, convective systems, and supercells of opposite directions of rotation though these were rare (found in less than 10% of cases).
Bar charts showing the relative frequencies of mergers, collisions, and splits in the 15 min preceding (left) significant tornadoes and (right) very large hail, as well as null events. Bars are colored by the type of interaction. Percentages are relative to the total number of cases in each sample.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
e. Prevalence of features in different environments
Now that we have shown what types of features are most common preceding tornadogenesis and large hail production, we investigate if they are more or less common in certain environments. In cases of significant tornadoes, these features varied in prevalence with low-level shear and CAPE (Fig. 10). A normalized probability density function is used to examine not only the frequency of each type but also its frequency relative to other types in the broader parameter space. Neighboring cells and supercells were common throughout this parameter space, especially where low-level shear was strong and CAPE was moderate. Boundaries and convective systems generally occupied opposite ends of this parameter space. Where shear was strongest and CAPE was smallest, convective systems became more common relative to other features. On the other hand, as shear became weaker and CAPE became larger, tornadoes were increasingly common near boundaries [consistent with Garner (2012) and Boustead et al. (2013)]. Although boundary vorticity may compensate for weak shear (Markowski et al. 1998b; Rasmussen et al. 2000; Coniglio and Jewell 2022), this does not explain why boundaries became less common as shear increased. Rather, this may reflect that stronger shear can deform these boundaries (especially drylines) into horizontal convective rolls (Ziegler et al. 1997; Peckham and Wicker 2000; Xue and Martin 2006; Rye and Duda 2007) that would not be detectable under this study’s methodology, or something else that this study cannot quantify. The separation between the distributions of boundaries and convective systems was most distinguishable when using CAPE and low-level shear, rather than ML3CAPE or a deeper-layer shear (not shown).
(left) Probability density function of the frequency of each type of feature in the 15 min preceding a significant tornado, per (top) 0–500-m BWD (m s−1) and (bottom) CAPE (J kg−1). To display the frequencies of each feature compared to others, the PDF of each type is normalized by the fraction of cases associated with that feature. As a result, the probabilities in each feature’s PDF add up not to 1.0 but to the fraction of cases associated with that feature. The probabilities of all features’ PDFs combined add up to 1.0. This way, the probabilities between different features can be compared such that equal probabilities mean that they are equally common in a given environment. (right) Mean environmental vertical profiles for all cases of significant tornadoes associated with (top) boundaries and (bottom) convective systems in the 15-min preceding tornadogenesis. The skew T is composed of the dewpoint mean (blue line) and interquartile range (IQR, blue shading), temperature mean (red line), and IQR (red shading), wet-bulb temperature mean (light blue solid line), and the MU parcel path (gray dashed line), LCL (blue horizontal mark), LFC (orange horizontal mark), and downdraft parcel path [as in downward CAPE (DCAPE), light blue dashed line] calculated from this mean profile. The hodograph consists of the mean 0–1-km (purple), 1–3-km (red), 3–6-km (orange), 6–9-km (gold), and 9–12-km (light yellow) shear vectors, and Bunkers RM storm motion (Bunkers et al. 2000, red dot) calculated from this mean profile, with 10 and 20 m s−1 storm-relative wind range rings. The mean hodograph was created by rotating each wind profile such that the 0–3-km BWD vector pointed east, then translating to a 0-kt (1 kt ≈ 0.51 m s−1) surface wind, as in Nixon and Allen (2022), for reasons discussed therein. This mean hodograph is then displayed in the storm-relative sense by subtracting the storm motion. All wind profiles are displayed in light gray, and ellipses mark the 90% confidence on the mean of the u and υ wind at each level. An assortment of parameters (their definitions and units defined in the data and methods section) are plotted on the left-hand side of each panel and calculated using the mean profile. Height is in meters, temperature is in degrees Celsius, and wind is in meters per second. The size of each sample is given in the top-right corner of each plot. “None” refers to cases that did not feature any interactions in this 15-min window.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
The difference in mean vertical profiles between boundary cases and convective system cases is also shown in Fig. 10. The difference in both 0–500-m shear and MUCAPE between samples was statistically significant at the 1% level using a Mann–Whitney U test (Mann and Whitney 1947). We remind the reader that some of these differences (especially in the thermodynamic profile) may also be tied to regional differences in environmental climatology; for instance, that the boundary tornado case appears similar to the hail cases that follow may be a function of their both being concentrated in the Great Plains. We also remind the reader that obtaining representative low-level shear and thermodynamic profiles near boundaries depends on modeling the accurate location of the boundary in the first place (Allen and Karoly 2014), and we cannot be fully confident that profiles were taken from the correct side of the boundary in each case (or that the boundary was even modeled). That said, the mean boundary profile is consistent with that found by Boustead et al. (2013), who used a reanalysis with smaller grid spacing.
The features preceding very large hail production varied less notably with these environmental parameters (Fig. 11). That said, similarly to tornado cases, boundaries became more common relative to other features when low-level shear was weaker, and CAPE was larger. Also similarly, these boundaries became especially numerous as this shear approached zero. As low-level shear increased, neighboring cells and convective systems became increasingly common relative to boundaries. That said, although convective systems rarely produced significant tornadoes when low-level shear was weak, such systems commonly produced hail in these environments.
As in Fig. 10, but for very large hail reports.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
The difference in mean vertical profiles between boundary cases and convective system cases is also shown in Fig. 11. Neither differences in CAPE nor low-level shear were statistically significant. We note that while these vertical profiles may provide useful prestorm, far-field environmental information, they are likely not representative of the inflow air mass in hail cases. As shown above, hailstorms are most commonly found in stable air masses north of boundaries and behind the gust fronts of convective systems. Thus, the true near-inflow profiles of hailstorms likely had more stable low-level thermodynamic profiles, if not near-surface inversions.
f. Prevalence of neighboring cells in different environments
Neighboring cells were common in cases of significant hazards, especially tornadoes, and particularly in the 10 min prior to hazard production. Given that these cells can impact storms differently as a function of the strengths of their outflows (Jewett et al. 2006; Hastings and Richardson 2016; Fischer and Dahl 2023) or the storm-relative inflow (Fischer and Dahl 2023), we now examine the prevalence of these cells in different environments. We consider that the LCL height may serve as a proxy for potential downdraft strength (Markowski et al. 2002; Murdzek et al. 2022) by way of enhanced evaporative cooling such that cells in higher LCL environments may have potentially stronger outflows. We also consider the mean 0–500-m storm-relative inflow [calculated using an observed storm motion for all tornadic cases and the Bunkers storm motion estimate for all other cases (Bunkers et al. 2000)]. For this section only, cases are excluded if they were found to be “anchored” on a boundary, or on the less unstable side of a boundary or convective system, so that storm-relative inflow estimates may be more realistic.
Rear-flank cells were much more prevalent with lower LCLs in the tornadic sample (Fig. 12). The difference in frequency of these rear-flank cells with LCL height in this sample is especially large relative to the nontornadic sample. For the tornadic cases, the favored position of these cells was closer, and more directly opposite the storm-relative inflow, with lower LCLs (with a density peak at 180°, as opposed to 198° for the high-LCL sample). These cells also remained closer to the storm for a longer period of time with lower LCLs (Fig. S31). Thus, as the environmental potential for stronger downdrafts decreased, neighboring cells were more commonly found near tornadic storms and were closer to them for a longer period of time. For the very large hail sample, storms were similarly accompanied by more left-/rear-flank cells with lower LCLs but were particularly free of rear-flank cells with higher LCLs (Fig. S33). We note that LCL height may be correlated with CIN such that lower LCLs may simply be accompanied by more neighboring cells because they imply better odds of convection initiation. However, upon a test of cell locations with MLCIN (Fig. S34), we find little difference in the frequency and position of nearby cells, especially with tornadoes. Therefore, we hypothesize that the increase in the frequency of nearby cells in lower LCLs has some physical significance and is not just a coincidence.
As in Fig. 7, but for the density of cells in two different samples subset by MLLCL (m) for tornadic cases, 10-min preceding hazard production. The size of each sample is given in the bottom right corner of each plot. The ERA5 mean 0–500-m storm-relative wind (using the Bunkers right vector for null cases and radar-estimated vector for all other cases) is plotted as a wind barb (knots).
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
We also find that rear-flank cells differ with changes in mean 0–500-m storm-relative inflow magnitude but mostly in their position rather than their frequency (Fig. 13). For the tornadic sample, when storm-relative inflow was stronger, the favorable positions of rear-flank cells exhibited less spread and were at an angle more directly opposite the storm-relative inflow vector (with a density peak at 178°, as opposed to 210° for the weak storm-relative inflow sample). In particularly strong storm-relative inflow, more rear-flank cells accompanied the tornadic storms for a longer period of time than they did nontornadic storms (Fig. S32). There was less of a difference in the positions of rear-flank cells in very large hail cases, but cases in weaker storm-relative inflow generally had more rear-flank cells, while cases in stronger inflow had fewer (Fig. S33).
As in Fig. 12, but for the density of cells in two different samples subset by 0–500-m storm-relative wind.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
4. Summary and discussion
In the half hour leading up to significant tornadoes and very large hail, their parent storms have no shortage of nearby features with which they could be interacting. In this sample of 476 cases, the most common features accompanying these storms were as follows:
Significant tornadoes:
-
Cells or supercells in the storm’s left/rear flank (but not its forward/right flank)
-
Convective systems, within which the storm was at the head, or embedded
Very large hail:
-
Boundaries, whereupon the parent storm was within relatively stable air
-
Cells or supercells in the storm’s forward/left flanks
-
Convective systems, within which the storm was at the tail
Although traditional practice in severe weather forecasting puts emphasis on discrete supercells, supercells that produced significant hazards without one of these nearby features were the exception rather than the rule, accounting for less than 10% of cases in this sample. This is important because our conceptual understanding of supercells and their hazards is based largely on the observation of supercells themselves, the simulation of supercells in isolation, and the knowledge of background environmental parameters. This study serves as a reminder that a storm’s environment is not limited to our parameterization of the vertical profile and that supercells are not always in isolation. Rather, they may be surrounded by a variety of features that may add new perspectives to our long-held conceptual understanding.
Although significant tornadoes can happen in a rather wide variety of background environmental parameters, their parent supercells were accompanied by a comparatively specific configuration of near-storm features. Regardless of the environment, significantly tornadic supercells were distinctively free of other cells in their forward flanks. This is similar to what is seen in cases of simulated supercells, whereupon the downdrafts from the storm itself can obstruct its access to warm, buoyant inflow air, consequently thwarting tornadogenesis (Gray and Frame 2021). Similarly, in this sample, a lack of strong convection in the forward flanks of supercells appears to be a condition for significant tornadoes.
The parent supercells of significant tornadoes were frequently accompanied by other cells in their rear flanks. This corroborates a similar observational study by Broyles et al. (2022), who hypothesized that these cells could be important for tornadogenesis. That these results mirror theirs is both a testament to similar/overlapping samples and also similar methodology. Others have found through simulation that these cells can strengthen a supercell’s outflow with their own downdrafts (Hastings and Richardson 2016; Fischer and Dahl 2023). Fischer and Dahl (2023) suggested that a supercell’s own downdraft may not be enough to instigate tornadogenesis in some environments and that the external downdrafts from other cells may be important. Consequently, our study finds that most tornadic storms are accompanied by cells in their rear flanks. Furthermore, the prevalence of these rear-flank cells varies with LCL height, which modulates evaporative cooling and potential downdraft strength in supercells (Markowski et al. 2002; Murdzek et al. 2022). In environments with low LCLs, where strong outflow is discouraged, we find that tornadic storms are accompanied by considerably more rear-flank cells, which remain near the storm for a longer period of time. This may just be because nearby cells, especially those producing 30-dBZ reflectivity on the lowest-tilt scan, are more likely with more saturated boundary layers. That said, we also hypothesize that these cells may be increasingly important in these environments to compensate for the otherwise weak downdraft potential given thermodynamics.
The effectiveness of rear-flank cells in instigating tornadogenesis may also depend on the low-level storm-relative winds. Excessive low-level storm-relative inflow has been noted to advect pretornadic vortices past an updraft too quickly (Fischer and Dahl 2023) and induce unfavorably strong updraft tilting (Peters et al. 2023), thus minimizing their opportunity to be amplified to tornadic strength. Consequently, Fischer and Dahl (2023) found that in these cases, this inflow must be countered by sufficiently strong outflow (especially from an external cell) to decelerate the progression of vortices and allow them more residence time beneath the updraft to amplify. This inflow/outflow “balance” has been noted in observed supercells to enhance tornado potential and maintenance (Dowell and Bluestein 2002). We found that the rear-flank cells themselves were neither more nor less common in stronger storm-relative inflow. However, the favorable positions of these cells became increasingly confined to a position closer to the storm and more directly opposite the storm-relative inflow as its magnitude increased. We also found that more cases of strong tornadoes are produced by convective systems (which presumably induce stronger outflow) in these cases. Since stronger outflow is necessary to maintain an inflow/outflow balance when storm-relative inflow is stronger, we postulate that stronger and more favorably placed cells may be necessary to support tornadogenesis in these environments. An example of common features noted before tornadogenesis is shown in Fig. 14.
An example of common features observed 10 min before a significant tornado. Shown is base reflectivity data (dBZ) from the Atlanta, GA (KFFC), WSR-88D on 28 Apr 2011, plotted using GRLevel2 (an EF3 tornado began 10 min later). The storm’s mesocyclone is marked with a black triangle, and its 30-dBZ precipitation footprint is circled in solid white. The storm’s forward, rear, right, and left flanks are labeled and bounded by dashed black lines. All neighboring cells (their 30-dBZ precipitation footprints) that reached 30-dBZ reflectivity for longer than 10 min are circled. Their directions of motion are denoted by white arrows. The approximate 0–500-m mean storm-relative wind for this case is drawn as a wind barb, estimated using the nearest observed raob sounding. The tornadic storm is the rightmost of the pair of supercells with respect to its orientation. It has an abundance of unorganized convection in its rear flank. With the exception of a small cell in its right flank that soon becomes a merger, the storm’s forward and right flanks are free of convection.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Although significant hail events tended to occur in a rather specific set of background environmental parameters in this sample, their parent supercells were accompanied by a rather wide variety of external features and interactions. Many storms were near boundaries, particularly on the stable side, as has been found in past studies (Schemm et al. 2016; Kunz et al. 2020; Magee and Davenport 2020). Otherwise, they were frequently accompanied by external storms, most typically at the tail end of another supercell or a convective system, and thus presumably subjected to more stable air from downdrafts in their forward flanks. That large hail growth is favored in cases of weaker low-level buoyancy corroborates Lin and Kumjian (2022); they found that beyond a certain magnitude, stronger low-level buoyancy can induce excessively strong low-level updrafts that shorten the residence time of hail embryos within the hail growth zone, thus stunting large hail growth. While stronger low-level storm-relative winds can also shorten this residence time (Dennis and Kumjian 2017; Kumjian et al. 2021; Lin and Kumjian 2022), Nixon et al. (2023) found evidence that their effect may be dampened when low-level buoyancy was weaker. Thus, we hypothesize that more stable air in the inflow of hailstorms may make hail production more likely. In contrast to tornadic cases, rear-flank cells were much less common in cases of very large hail. This was reflected both in the analysis of neighboring cells and in that most hailstorms were positioned at the tail end of a system or pair of supercells. Unfortunately, further analysis into the positions of nearby cells relative to the storm-relative inflow was not possible here since the sample size of hailstorms not in a modified outflow or postfrontal air mass (where ERA5 proximity soundings may not be representative) was small. However, further work may be necessary to understand if this lack of rear-flank cells is significant for hail production. An example of common features noted before very large hail reports is shown in Fig. 15.
As in Fig. 14, but for common features observed 10 min before a very large hail report. Shown is base reflectivity data (dBZ) from the Brownsville, TX (KBRO), WSR-88D on 3 Apr 2017 (3 in. hail was reported 10 min later). The hailstorm is the tail of two supercells, which are both behind a fine-line boundary signature; with knowledge of both of these, this storm is likely elevated above a stable boundary layer. It has a couple strong cells nearby that move quickly out of its rear flank.
Citation: Weather and Forecasting 39, 10; 10.1175/WAF-D-23-0117.1
Along with past studies (e.g., Rogers 2012; Hastings and Richardson 2016; Magee and Davenport 2020; Broyles et al. 2022; Flournoy et al. 2022; Fischer and Dahl 2023; Fischer et al. 2023; Lyza and Flournoy 2023), we find evidence that storm interactions affect the future state of a storm and may discriminate between hazardous and nonhazardous storms even up to 30 min in advance and 50 km away. Thus, a better understanding of these interactions may be important to the improvement of warnings, which are presently based primarily on a storm’s background environment and current state (Donavon and Jungbluth 2007; Blair et al. 2011; Smith et al. 2015; Gibbs and Bowers 2019). These results advocate for a variety of future studies that may better observe or model storms’ interaction with these external features. Although cell “mergers” or “collisions” are referenced frequently in operations, we found that hazardous and nonhazardous storms may also be distinguished by cells that do not merge or collide but simply remain near the storm. At what distance and lead time can these other cells impact a storm? What is the difference in impacts between cells that merge, collide, or just remain near the storm for a prolonged duration? How might the impacts change with the storm-relative speed of a collision?
This study presents a manual approach to assessing potential storm interactions by a human observer (such that could be used in operations). Such a manual approach can introduce a degree of subjectivity in analysis, and consequently, different people may assess features differently. To minimize this subjectivity, future studies could reproduce this analysis using more human observers (e.g., Flournoy et al. 2022; Fischer and Dahl 2023) or computers. Larger samples may also be desired and obtainable. Future work may be required to better understand how these features relate to the background environment and if they become more or less necessary for tornadogenesis and hail production depending on the environment. This could be achieved through more fine-scale in situ observations of storms and complementary high-resolution, storm-scale numerical simulations.
Although the relationship between a supercell and its environment has been firmly established in operations, its relationship with external interactions appears similarly important. These interactions can be complex, and the impact of any particular external feature on a given storm may never be fully anticipated. That said, with this analysis of external features and their potential impacts, we hope to build a fuller conceptual model of supercell hazard production that can assist both forecasting and warning decisions.
For the storms that did not have identifiable mesocyclones, this criterion was omitted.
This analysis was not done for the storms that did not have identifiable mesocyclones.
Acknowledgments.
We would like to acknowledge Matthew Kumjian and Chris Broyles for their feedback and guidance. C. J. Nixon and J. T. Allen were supported by the National Science Foundation under Grant AGS-1855054. C. J. Nixon was also supported by the Earth and Ecosystem Science Ph.D. program at Central Michigan University. M. Taszarek was supported by a grant from the Polish National Science Center (2020/39/D/ST10/00768).
Data availability statement.
Data used in this paper were derived from ERA5 reanalysis (openly available through the Climate Data Source at https://cds.climate.copernicus.eu/#!/home) and storm reports from the Storm Prediction Center (https://www.spc.noaa.gov/wcm/). The authors’ dataset of quality-controlled cases with corresponding notes of observed storm interactions is available through Zenodo (https://doi.org/10.5281/zenodo.10672225).
REFERENCES
Adlerman, E. J., K. K. Droegemeier, and R. Davies-Jones, 1999: A numerical simulation of cyclic mesocyclogenesis. J. Atmos. Sci., 56, 2045–2069, https://doi.org/10.1175/1520-0469(1999)056<2045:ANSOCM>2.0.CO;2.
Allen, J. T., and D. J. Karoly, 2014: A climatology of Australian severe thunderstorm environments 1979–2011: Inter-annual variability and ENSO influence. J. Climate, 34, 81–97, https://doi.org/10.1002/joc.3667.
Allen, J. T., and M. K. Tippett, 2015: The characteristics of United States hail reports: 1955-2014. Electron. J. Severe Storms Meteor., 10, 1–31, https://doi.org/10.55599/ejssm.v10i3.60.
Allen, J. T., D. J. Karoly, and K. J. Walsh, 2014: Future Australian severe thunderstorm environments. Part I: A novel evaluation and climatology of convective parameters from two climate models for the late twentieth century. J. Climate, 27, 3827–3847, https://doi.org/10.1175/JCLI-D-13-00425.1.
Allen, J. T., M. K. Tippett, Y. Kaheil, A. H. Sobel, C. Lepore, S. Nong, and A. Muehlbauer, 2017: An extreme value model for U.S. hail size. Mon. Wea. Rev., 145, 4501–4519, https://doi.org/10.1175/MWR-D-17-0119.1.
Allen, J. T., I. M. Giammanco, M. R. Kumjian, H. J. Punge, Q. Zhang, P. Groenemeijer, M. Kunz, and K. Ortega, 2020: Understanding hail in the Earth system. Rev. Geophys., 58, e2019RG000665, https://doi.org/10.1029/2019RG000665.
Ashley, W. S., 2007: Spatial and temporal analysis of tornado fatalities in the United States: 1880–2005. Wea. Forecasting, 22, 1214–1228, https://doi.org/10.1175/2007WAF2007004.1.
Ashley, W. S., A. M. Haberlie, and J. Strohm, 2019: A climatology of quasi-linear convective systems and their hazards in the United States. Wea. Forecasting, 34, 1605–1631, https://doi.org/10.1175/WAF-D-19-0014.1.
Atkins, N. T., and M. St. Laurent, 2009: Bow echo mesovortices. Part II: Their genesis. Mon. Wea. Rev., 137, 1514–1532, https://doi.org/10.1175/2008MWR2650.1.
Atkins, N. T., M. L. Weisman, and L. J. Wicker, 1999: The influence of preexisting boundaries on supercell evolution. Mon. Wea. Rev., 127, 2910–2927, https://doi.org/10.1175/1520-0493(1999)127<2910:TIOPBO>2.0.CO;2.
Banacos, P. C., and D. M. Schultz, 2005: The use of moisture flux convergence in forecasting convective initiation: Historical and operational perspectives. Wea. Forecasting, 20, 351–366, https://doi.org/10.1175/WAF858.1.
Barker, L. J., III, 2006: A potentially valuable WSR-88D severe storm pre-cursor signature in highly dynamic, low cape, high shear environments. Preprints, 23rd Conf. on Severe Local Storms, St. Louis, MO, Amer. Meteor. Soc., 3.6, https://www.weather.gov/media/ilx/Research-Case_Studies/barker_sls06.pdf.
Beveridge, S. L., J. L. Houser, and S. R. Marzola, 2019: A statistical evaluation of tornado-production tendencies of southernmost supercells compared to adjacent supercells in a north–south-oriented line. Electron. J. Severe Storms Meteor., 14, 1–33, https://doi.org/10.55599/ejssm.v14i1.71.
Blair, S. F., and J. W. Leighton, 2012: Creating high-resolution hail datasets using social media and post-storm ground surveys. Electron. J. Oper. Meteor., 13, 32–45.
Blair, S. F., D. R. Deroche, J. M. Boustead, J. W. Leighton, B. L. Barjenbruch, and W. P. Gargan, 2011: A radar-based assessment of the detectability of giant hail. Electron. J. Severe Storms Meteor., 6, 1–30, https://doi.org/10.55599/ejssm.v6i7.34.
Blair, S. F., and Coauthors, 2017: High-resolution hail observations: Implications for NWS warning operations. Wea. Forecasting, 32, 1101–1119, https://doi.org/10.1175/WAF-D-16-0203.1.
Boustead, J. M., B. E. Mayes, W. Gargan, J. L. Leighton, G. Phillips, and P. N. Schumacher, 2013: Discriminating environmental conditions for significant warm sector and boundary tornadoes in parts of the Great Plains. Wea. Forecasting, 28, 1498–1523, https://doi.org/10.1175/WAF-D-12-00102.1.
Branick, M. L., 1996: A comprehensive glossary of weather terms for storm spotters. NOAA Tech. Memo. NWS SR-145, 28 pp., https://repository.library.noaa.gov/view/noaa/7223.
Brooks, H. E., and J. Correia Jr., 2018: Long-term performance metrics for National Weather Service tornado warnings. Wea. Forecasting, 33, 1501–1511, https://doi.org/10.1175/WAF-D-18-0120.1.
Brown, R. A., L. R. Lemon, and D. W. Burgess, 1978: Tornado detection by pulsed Doppler radar. Mon. Wea. Rev., 106, 29–38, https://doi.org/10.1175/1520-0493(1978)106<0029:TDBPDR>2.0.CO;2.
Browning, K. A., and G. A. Monk, 1982: A simple model for the synoptic analysis of cold fronts. Quart. J. Roy. Meteor. Soc., 108, 435–452, https://doi.org/10.1002/qj.49710845609.
Broyles, J. C., C. K. Potvin, G. L. Dial, J. Murnan, S. Shores, A. D. Lyons, M. S. Elliott, and A. R. Cook, 2022: Tornadogenesis in high-end tornadic supercells (Part 1) – The RFD surge, RFD occlusion and cell mergers. 30th Conf. on Severe Local Storms, Santa Fe, NM, Amer. Meteor. Soc., 13, https://www.spc.noaa.gov/publications/broyles/tgen1.pdf.
Bunkers, M. J., B. A. Klimowski, J. W. Zeitler, R. L. Thompson, and M. L. Weisman, 2000: Predicting supercell motion using a new hodograph technique. Wea. Forecasting, 15, 61–79, https://doi.org/10.1175/1520-0434(2000)015<0061:PSMUAN>2.0.CO;2.
Caruso, J. M., and J. M. Davies, 2005: Tornadoes in nonmesocyclone environments with pre-existing vertical vorticity along convergence boundaries. Electron. J. Oper. Meteor., 6, 1–36.
Chasteen, M. B., and S. E. Koch, 2022: Multiscale aspects of the 26–27 April 2011 tornado outbreak. Part I: Outbreak chronology and environmental evolution. Mon. Wea. Rev., 150, 309–335, https://doi.org/10.1175/MWR-D-21-0013.1.
Christian, T. W., and R. M. Wakimoto, 1989: The relationship between radar reflectivities and clouds associated with horizontal roll convection on 8 August 1982. Mon. Wea. Rev., 117, 1530–1544, https://doi.org/10.1175/1520-0493(1989)117<1530:TRBRRA>2.0.CO;2.
Coffer, B. E., M. D. Parker, R. L. Thompson, B. T. Smith, and R. E. Jewell, 2019: Using near-ground storm relative helicity in supercell tornado forecasting. Wea. Forecasting, 34, 1417–1435, https://doi.org/10.1175/WAF-D-19-0115.1.
Coffer, B. E., M. Taszarek, and M. D. Parker, 2020: Near-ground wind profiles of tornadic and nontornadic environments in the United States and Europe from ERA5 reanalyses. Wea. Forecasting, 35, 2621–2638, https://doi.org/10.1175/WAF-D-20-0153.1.
Concannon, P. R., H. E. Brooks, and C. A. Doswell, 2000: Climatological risk of strong and violent tornadoes in the United States. Preprints, Second Symp. on Environmental Applications, Long Beach, CA, Amer. Meteor. Soc., 212–219.
Coniglio, M. C., and M. D. Parker, 2020: Insights into supercells and their environments from three decades of targeted radiosonde observations. Mon. Wea. Rev., 148, 4893–4915, https://doi.org/10.1175/MWR-D-20-0105.1.
Coniglio, M. C., and R. E. Jewell, 2022: SPC mesoscale analysis compared to field-project soundings: Implications for supercell environment studies. Mon. Wea. Rev., 150, 567–588, https://doi.org/10.1175/MWR-D-21-0222.1.
Corfidi, S. F., S. J. Corfidi, and D. M. Schultz, 2008: Elevated convection and castellanus: Ambiguities, significance, and questions. Wea. Forecasting, 23, 1280–1303, https://doi.org/10.1175/2008WAF2222118.1.
Crum, T. D., and R. L. Alberty, 1993: The WSR-88D and the WSR-88D operational support facility. Bull. Amer. Meteor. Soc., 74, 1669–1688, https://doi.org/10.1175/1520-0477(1993)074<1669:TWATWO>2.0.CO;2.
Czernecki, B., M. Taszarek, M. Marosz, M. Półrolniczak, L. Kolendowicz, A. Wyszogrodzki, and J. Szturc, 2019: Application of machine learning to large hail prediction - The importance of radar reflectivity, lightning occurrence and convective parameters derived from ERA5. Atmos. Res., 227, 249–262, https://doi.org/10.1016/j.atmosres.2019.05.010.
Davies, J. M.,