Many studies have aimed to identify novel storm characteristics that are indicative of current or future severe weather potential using a combination of ground-based radar observations and severe reports. However, this is often done on a small scale using limited case studies on the order of tens to hundreds of storms due to how time-intensive this process is. Herein, we introduce the GridRad-Severe dataset, a database including ~100 severe weather days per year and upwards of 1.3 million objectively tracked storms from 2010-2019. Composite radar volumes spanning objectively determined, report-centered domains are created for each selected day using the GridRad compositing technique, with dates objectively determined using report thresholds defined to capture the highest-end severe weather days from each year, evenly distributed across all severe report types (tornadoes, severe hail, and severe wind). Spatiotemporal domain bounds for each event are objectively determined to encompass both the majority of reports as well as the time of convection initiation. Severe weather reports are matched to storms that are objectively tracked using the radar data, so the evolution of the storm cells and their severe weather production can be evaluated. Herein, we apply storm mode (single cell, multicell, or mesoscale convective system) and right-moving supercell classification techniques to the dataset, and revisit various questions about severe storms and their bulk characteristics posed and evaluated in past work. Additional applications of this dataset are reviewed for possible future studies.

Given this large dataset of severe storms, questions about storm structure of very specific storm types can be investigated using what is still a large subsample of the total GridRad-Severe dataset. This study compares populations of tornadic non-supercellular MCS storm cells to their nontornadic counterparts, focusing on nontornadic storms that have similar radar characteristics to tornadic storms. Comparison of single-polarization radar variables during storm lifetimes show that median values of low-level, mid-level, and column-maximum azimuthal shear, as well as low-level radial divergence, enable the highest degree of separation between tornadic and nontornadic storms. Focusing on low-level azimuthal shear values, null storms were randomly selected such that the distribution of null low-level azimuthal shear values matches the distribution of tornadic values. After isolating the null cases from the nontornadic population, signatures emerge in single-polarization data that enable discrimination between nontornadic and tornadic storms. In comparison, dual-polarization variables show little deviation between storm types. Tornadic storms both at tornadogenesis and at 20-minute lead time show collocation of the primary storm updraft with enhanced near-surface rotation and convergence, facilitating the non-mesocyclonic tornadogenesis processes.

With this additional knowledge about the structure of tornadic vs. nontornadic storms and which radar variables best differentiate the two, machine learning methods can be used to learn the differences between these storm type at various lead times and improve tornado predictability. A convolutional neural network was trained on tornadic and nontornadic data where the nontornadic data were either sampled from storms that have similar radar characteristics to tornadic storms as in the PMM analyses or sampled from the entire population of non-supercellular MCS storms. These models were then tested on independent data from 2020-2021, again either including all tornadic storms and sampling nontornadic cases as in the PMM analyses or including all tornadic and nontornadic storms. Models that were tested on all tornadic and nontornadic storms, whether they were trained and validated on datasets including sampled strong nontornadic storms or a sample of all nontornadic storms, both performed well below the baseline performance metrics from the NWS. However, when the model was trained, validated, and tested using samples of all tornadic storms and only strong nontornadic storms, model test performance far exceeded the baseline NWS metrics. Performance metrics include a probability of detection (POD) of 79%, a false alarm ratio (FAR) of 58%, and a CSI of 0.38. Compared to the NWS metrics of 49%, 75%, and 0.2, respectively, this model shows clear promise as a supplemental forecasting tool for scenarios where a storm is identified as (at least) borderline tornadic. However, further analyses of the model performance scaled to account for the true proportion of tornadic vs. nontornadic storms shows that it was the unnatural ratio of tornadic to nontornadic storms, and not the focus on strong nontornadic storms, that was the cause for the improved model performance.

Finally, a brief analysis of the underlying populations and their demographic characteristics in the vicinity of tornadoes are examined. Special attention is given to non-supercellular MCS storms, as well as discrete supercells, whose tornadoes are often a main focus of tornado research in the U.S. Analyses show that groups making up ~3% or less of the CONUS mean population typically have lower relative population densities in the vicinity of storms. The Black or African American Alone demographic has higher relative populations in the vicinity of all tornadoes compared to their CONUS mean population density, as do all Non-Hispanic categories (Not Hispanic, Non-Hispanic White and Non-Hispanic Black). Comparing population densities near specific types of tornadoes (i.e., mode and combination of mode and human impact) to their densities near all tornadoes, the White Alone demographic has population densities near the CONUS mean for supercellular tornadoes, but that density jumps 6-7 percentage points in the vicinity of deadly supercellular tornadoes when examining underlying population density by deadly event and by death, suggesting that the deadliest supercellular tornadoes occur in predominantly White areas. On average, populations in the vicinity of all tornadoes have ~75-80% higher Black or African American Alone and Non-Hispanic Black densities when compared to the CONUS mean, with those demographics' relative densities only increasing when isolating MCS tornadoes and deadly MCS tornadoes, suggesting that the deadliest MCS tornadoes preferentially occur in areas with relatively higher Black or African American Alone and Non-Hispanic Black populations. One particularly striking result is that the mean Social Vulnerability Index (SVI) of populations near all tornadoes is just barely above the CONUS mean (0.52 vs. CONUS mean of 0.51), but is slightly lower for supercellular tornadoes (0.49) and higher for MCS tornadoes (0.57). Therefore, MCS tornadoes tend to occur in areas that are less resilient to natural disasters than both the CONUS mean and areas in the vicinity of supercellular tornadoes. For both MCS and supercellular tornadoes that were associated with deaths or injuries, the local SVI is higher, likely pointing to the applicability of SVI in identifying areas less resilient to natural disasters.