Search

We present an overviewof recentwork on using artificial intelligence/machine learning techniques for forecasting convective weather and its associated hazards, including tornadoes, hail, wind, and lightning. These high-impact phenomena globally cause both massive property damage and loss of life yet they are quite challenging to forecast. Given the recent explosion in developing machine learning techniques across the weather spectrum and the fact that the skillful prediction of convective weather has immediate societal benefits, we present a thorough review of the current state of the art in artificial intelligence and machine learning techniques for convective hazards. Our review includes both traditional approaches, including support vector machines and decision trees as well as deep learning approaches. We highlight the challenges in developing machine learning approaches to forecast these phenomena across a variety of spatial and temporal scales. We end with a discussion of promising areas of future work for ML for convective weather, including a discussion of the need to create trustworthy AI forecasts that can be used for forecasters in real-time and the need for active cross-sector collaboration on testbeds to validate machine learning methods in operational situations.

Accurate vertical velocity retrieval from dual-Doppler analysis (DDA) is a long-standing problem of radar meteorology. Typical radar scanning strategies poorly observe the vertical component of motion, leading to large uncertainty in vertical velocity estimates. Using a vertical vorticity equation constraint in addition to a mass conservation constraint in DDA has shown promise in improving vertical velocity retrievals. However, observation system simulation experiments (OSSEs) suggest this technique requires rapid radar volume scans to realize the improvements due to the vorticity tendency term in the vertical vorticity constraint. Here, the vertical vorticity constraint DDA is tested with real, rapid-scan radar data to validate prior OSSEs results. Generally, the vertical vorticity constraint DDA produced more accurate vertical velocities from DDAs than those that did not use the constraint. When the time between volume scans was greater than 30 s, the vertical velocity accuracy was significantly affected by the vorticity tendency estimation method. A technique that uses advection correction on provisional DDA wind fields to shorten the discretization interval for the vorticity tendency calculation improved the vertical velocity retrievals for longer times between volume scans. The skill of these DDAs was similar to those using a shorter time between volume scans. These improvements were due to increased accuracy of the vertical vorticity tendency using the advection correction technique. The real radar data tests also revealed that the vertical vorticity constraint DDAs are more forgiving to radar data errors. These results suggest that vertical vorticity constraint DDA with rapid-scan radars should be prioritized for kinematic analyses.

Storm displacement errors can arise from a number of potential sources of error within a data assimilation (DA) and forecast system. Conversely, storm displacement errors can cause issues for storm-scale, ensemble-based systems using an ensemble Kalman filter (EnKF), such as NSSL’s Warn-on-Forecast System (WoFS). A previous study developed a fully grid-based feature alignment technique (FAT) to mitigate these phase errors and their impacts. However, that study developed and tested the FAT for single-storm cases. This study advances that work by implementing an object-based merging and matching technique into the FAT and tests the updated FAT in more complex scenarios of multiple storms. Ensemble-based experiments are conducted with and without the FAT for each of the scenarios. The experiments’ analyses and forecasts of storm-related fields are then evaluated using subjective and objective methods. Results from these idealized multiple-storm experiments continue to reveal the potential benefits of correcting storm displacement errors. For example, running the FAT even once can mitigate the “spinup” period experienced by the no-FAT experiments. The new results also show that running the FAT prior to every DA cycling step generally leads to more skillful forecasts at the smaller scales, especially in earlier-initialized forecasts. However, repeatedly running the FAT prior to every DA step can eventually lead to deterioration in analyses and forecasts. Potential solutions to this problem include using longer cycling intervals and running the FAT prior to DA less often. Additional ways to improve the FAT along with other results are presented and discussed. Significance Statement The purpose of this work is to explore the impact of correcting storm displacements on analyses and forecasts of storms using an ensemble-based data assimilation and forecast system in an idealized framework. Storm displacement errors are a common problem in current operational and experimental storm-scale forecast systems, so understanding their impact on these systems and providing a method to help mitigate them is important. Results from this study indicate that correcting storm displacement errors with the feature alignment technique can greatly improve analyses and forecasts in multiple-storm scenarios. Future work will focus on exploring the impact of correcting storm displacement errors in a real-data, storm-scale data assimilation and forecast system.

Many tornadoes are unreported because of lack of observers or are underrated in intensity, width, or track length because of lack of damage indicators. These reporting biases substantially degrade estimates of tornado frequency and thereby undermine important endeavors such as studies of climate impacts on tornadoes and cost–benefit analyses of tornado damage mitigation. Building on previous studies, we use a Bayesian hierarchical modeling framework to estimate and correct for tornado reporting biases over the central United States during 1975–2018. The reporting biases are treated as a univariate function of population density. We assess how these biases vary with tornado intensity, width, and track length and over the analysis period. We find that the frequencies of tornadoes of all kinds, but especially stronger or wider tornadoes, have been substantially underestimated. Most strikingly, the Bayesian model estimates that there have been approximately 3 times as many tornadoes capable of (E)F2+ damage as have been recorded as (E)F2+ [(E)F indicates a rating on the (enhanced) Fujita scale]. The model estimates that total tornado frequency changed little over the analysis period. Statistically significant trends in frequency are found for tornadoes within certain ranges of intensity, pathlength, and width, but it is unclear what proportion of these trends arise from changes in damage survey practices. Simple analyses of the tornado database corroborate many of the inferences from the Bayesian model. Significance Statement Prior studies have shown that the probabilities of a tornado being reported and of its intensity, track length, and width being accurately estimated are strongly correlated with the local population density. We have developed a sophisticated statistical model that accounts for these population-dependent tornado reporting biases to improve estimates of tornado frequency in the central United States. The bias-corrected tornado frequency estimates differ markedly from the official tornado climatology and have important implications for tornado risk assessment, damage mitigation, and studies of climate change impacts on tornado activity.

Thunderstorm mode strongly impacts the likelihood and predictability of tornadoes and other hazards, and thus is of great interest to severe weather forecasters and researchers. It is often impossible for a forecaster to manually classify all the storms within convection-allowing model (CAM) output during a severe weather outbreak, or for a scientist to manually classify all storms in a large CAM or radar dataset in a timely manner. Automated storm classification techniques facilitate these tasks and provide objective inputs to operational tools, including machine learning models for predicting thunderstorm hazards. Accurate storm classification, however, requires accurate storm segmentation. Many storm segmentation techniques fail to distinguish between clustered storms, thereby missing intense cells, or to identify cells embedded within quasi-linear convective systems that can produce tornadoes and damaging winds. Therefore, we have developed an iterative technique that identifies these constituent storms in addition to traditionally identified storms. Identified storms are classified according to a seven-mode scheme designed for severe weather operations and research. The classification model is a hand-developed decision tree that operates on storm properties computed from composite reflectivity and midlevel rotation fields. These properties include geometrical attributes, whether the storm contains smaller storms or resides within a larger-scale complex, and whether strong rotation exists near the storm centroid. We evaluate the classification algorithm using expert labels of 400 storms simulated by the NSSL Warn-on-Forecast System or analyzed by the NSSL Multi-Radar/Multi-Sensor product suite. The classification algorithm emulates expert opinion reasonably well (e.g., 76% accuracy for supercells), and therefore could facilitate a wide range of operational and research applications. Significance Statement We have developed a new technique for automatically identifying intense thunderstorms in model and radar data and classifying storm mode, which informs forecasters about the risks of tornadoes and other high-impact weather. The technique identifies storms that are often missed by other methods, including cells embedded within storm clusters, and successfully classifies important storm modes that are generally not included in other schemes, such as rotating cells embedded within quasi-linear convective systems. We hope the technique will facilitate a variety of forecasting and research efforts.

The National Severe Storms Laboratory (NSSL) Warn-on-Forecast System (WoFS) is an experimental real-time rapidly updating convection-allowing ensemble that provides probabilistic short-term thunderstorm forecasts. This study evaluates the impacts of reducing the forecast model horizontal grid spacing Δx from 3 to 1.5 km on the WoFS deterministic and probabilistic forecast skill, using 11 case days selected from the 2020 NOAA Hazardous Weather Testbed (HWT) Spring Forecasting Experiment (SFE). Verification methods include (i) subjective forecaster impressions; (ii) a deterministic object-based technique that identifies forecast reflectivity and rotation track storm objects as contiguous local maxima in the composite reflectivity and updraft helicity fields, respectively, and matches them to observed storm objects; and (iii) a recently developed algorithm that matches observed mesocyclones to mesocyclone probability swath objects constructed from the full ensemble of rotation track objects. Reducing Δx fails to systematically improve deterministic skill in forecasting reflectivity object occurrence, as measured by critical success index (CSIDET), a metric that incorporates both probability of detection (PODDET) and false alarm ratio (FARDET). However, compared to the Δx = 3 km configuration, the Δx = 1.5 km WoFS shows improved midlevel mesocyclone detection, as evidenced by its statistically significant (i) higher CSIDET for deterministic midlevel rotation track objects and (ii) higher normalized area under the performance diagram curve (NAUPDC) score for probability swath objects. Comparison between Δx = 3 km and Δx = 1.5 km reflectivity object properties reveals that the latter have 30% stronger mean updraft speeds, 17% stronger median 80-m winds, 67% larger median hail diameter, and 28% higher median near-storm-maximum 0–3-km storm-relative helicity.

Tornadoes have Lorenzian predictability horizons O(10) min, and convection-allowing ensemble prediction systems (EPSs) often provide probabilistic guidance of such events to forecasters. Given the O(0.1)-km length scale of tornadoes and O(1)-km scale of mesocyclones, operational models running at horizontal grid spacings (Δx) of 3 km may not capture narrower mesocyclones (typical of the southeastern United States) and certainly do not resolve most tornadoes per se. In any case, it requires O(50) times more computer power to reduce Δx by a factor of 3. Herein, to determine value in such an investment, we compare two EPSs, differing only in Δx (3 vs 1 km), for four low-CAPE, high-shear cases. Verification was grouped as 1) deterministic, traditional methods using pointwise evaluation, 2) a scale-aware probabilistic metric, and 3) a novel method via object identification and information theory. Results suggest 1-km forecasts better detect storms and any associated rapid low- and midlevel rotation, but at the cost of weak–moderate reflectivity forecast skill. The nature of improvement was sensitive to the case, variable, forecast lead time, and magnitude, precluding a straightforward aggregation of results. However, the distribution of object-specific information gain over all cases consistently shows greater average benefit from the 1-km EPS. We also reiterate the importance of verification methodology appropriate for the hazard of interest.

Techniques to mitigate analysis errors arising from the nonsimultaneity of data collections typically use advection-correction procedures based on the hypothesis (frozen turbulence) that the analyzed field can be represented as a pattern of unchanging form in horizontal translation. It is more difficult to advection correct the radial velocity than the reflectivity because even if the vector velocity field satisfies this hypothesis, its radial component does not—but that component does satisfy a second-derivative condition. We treat the advection correction of the radial velocity (υ r ) as a variational problem in which errors in that second-derivative condition are minimized subject to smoothness constraints on spatially variable pattern-translation components (U, V). The Euler–Lagrange equations are derived, and an iterative trajectory-based solution is developed in which U, V, and υ r are analyzed together. The analysis code is first verified using analytical data, and then tested using Atmospheric Imaging Radar (AIR) data from a band of heavy rainfall on 4 September 2018 near El Reno, Oklahoma, and a decaying tornado on 27 May 2015 near Canadian, Texas. In both cases, the analyzed υ r field has smaller root-mean-square errors and larger correlation coefficients than in analyses based on persistence, linear time interpolation, or advection correction using constant U and V. As some experimentation is needed to obtain appropriate parameter values, the procedure is more suitable for non-real-time applications than use in an operational setting. In particular, the degree of spatial variability in U and V, and the associated errors in the analyzed υ r field are strongly dependent on a smoothness parameter.

An improved understanding of common differences between tornadic and nontornadic supercells is sought using a large set of observations from the operational NEXRAD WSR-88D polarimetric radar network in the contiguous United States. In particular, data from 478 nontornadic and 294 tornadic supercells during a 7-yr period (2011–17) are used to produce probability-matched composite means of microphysical and kinematic variables. Means, which are centered on echo-top maxima and in a horizontal coordinate system rotated such that storm motion points in the positive x dimension, are created in altitude relative to ground level at times of peak echo-top altitude and peak midlevel rotation for nontornadic supercells and times at and prior to the first tornado in tornadic supercells. Robust differences between supercell types are found, with consistent characteristics at and preceding tornadogenesis in tornadic storms. In particular, the mesocyclone is found to be vertically aligned in tornadic supercells and misaligned in nontornadic supercells. Microphysical differences found include a low-level radar reflectivity hook echo aligned with and ~10 km right of storm center in tornadic supercells and displaced 5–10 km down-motion in nontornadic supercells, a low-to-midlevel differential radar reflectivity dipole that is oriented more parallel to storm motion in tornadic supercells and more perpendicular in nontornadic supercells, and a separation between enhanced differential radar reflectivity and specific differential phase (with unique displacement-relative correlation coefficient reductions) at low levels that is more perpendicular to storm motion in tornadic supercells and more parallel in nontornadic supercells.

The NOAA Warn-on-Forecast System (WoFS) is an experimental rapidly updating convection-allowing ensemble designed to provide probabilistic operational guidance on high-impact thunderstorm hazards. The current WoFS uses physics diversity to help maintain ensemble spread. We assess the systematic impacts of the three WoFS PBL schemes—YSU, MYJ, and MYNN—using novel, object-based methods tailored to thunderstorms. Very short forecast lead times of 0–3 h are examined, which limits phase errors and thereby facilitates comparisons of observed and model storms that occurred in the same area at the same time. This evaluation framework facilitates assessment of systematic PBL scheme impacts on storms and storm environments. Forecasts using all three PBL schemes exhibit overly narrow ranges of surface temperature, dewpoint, and wind speed. The surface biases do not generally decrease at later forecast initialization times, indicating that systematic PBL scheme errors are not well mitigated by data assimilation. The YSU scheme exhibits the least bias of the three in surface temperature and moisture and in many sounding-derived convective variables. Interscheme environmental differences are similar both near and far from storms and qualitatively resemble the differences analyzed in previous studies. The YSU environments exhibit stronger mixing, as expected of nonlocal PBL schemes; are slightly less favorable for storm intensification; and produce correspondingly weaker storms than the MYJ and MYNN environments. On the other hand, systematic interscheme differences in storm morphology and storm location forecast skill are negligible. Overall, the results suggest that calibrating forecasts to correct for systematic differences between PBL schemes may modestly improve WoFS and other convection-allowing ensemble guidance at short lead times.

An object-based verification method for short-term, storm-scale probabilistic forecasts was developed and applied to mesocyclone guidance produced by the experimental Warn-on-Forecast System (WoFS) in 63 cases from 2017 to 2018. The probabilistic mesocyclone guidance was generated by calculating gridscale ensemble probabilities from WoFS forecasts of updraft helicity (UH) in layers 2–5 km (midlevel) and 0–2 km (low-level) above ground level (AGL) aggregated over 60-min periods. The resulting ensemble probability swaths are associated with individual thunderstorms and treated as objects with a single, representative probability value prescribed. A mesocyclone probability object, conceptually, is a region bounded by the ensemble forecast envelope of a mesocyclone track for a given thunderstorm over 1 h. The mesocyclone probability objects were matched against rotation track objects in Multi-Radar Multi-Sensor data using the total interest score, but with the maximum displacement varied between 0, 9, 15, and 30 km. Forecast accuracy and reliability were assessed at four different forecast lead time periods: 0–60, 30–90, 60–120, and 90–150 min. In the 0–60-min forecast period, the low-level UH probabilistic forecasts had a POD, FAR, and CSI of 0.46, 0.45, and 0.31, respectively, with a probability threshold of 22.2% (the threshold of maximum CSI). In the 90–150-min forecast period, the POD and CSI dropped to 0.39 and 0.27 while FAR remained relatively unchanged. Forecast probabilities > 60% overpredicted the likelihood of observed mesocyclones in the 0–60-min period; however, reliability improved when allowing larger maximum displacements for object matching and at longer lead times.

The 2016–18 NOAA Hazardous Weather Testbed (HWT) Spring Forecasting Experiments (SFE) featured the Community Leveraged Unified Ensemble (CLUE), a coordinated convection-allowing model (CAM) ensemble framework designed to provide empirical guidance for development of operational CAM systems. The 2017 CLUE included 81 members that all used 3-km horizontal grid spacing over the CONUS, enabling direct comparison of forecasts generated using different dynamical cores, physics schemes, and initialization procedures. This study uses forecasts from several of the 2017 CLUE members and one operational model to evaluate and compare CAM representation and next-day prediction of thunderstorms. The analysis utilizes existing techniques and novel, object-based techniques that distill important information about modeled and observed storms from many cases. The National Severe Storms Laboratory Multi-Radar Multi-Sensor product suite is used to verify model forecasts and climatologies of observed variables. Unobserved model fields are also examined to further illuminate important intermodel differences in storms and near-storm environments. No single model performed better than the others in all respects. However, there were many systematic intermodel and intercore differences in specific forecast metrics and model fields. Some of these differences can be confidently attributed to particular differences in model design. Model intercomparison studies similar to the one presented here are important to better understand the impacts of model and ensemble configurations on storm forecasts and to help optimize future operational CAM systems.

Observation system simulation experiments are used to evaluate different dual-Doppler analysis (DDA) methods for retrieving vertical velocity w at grid spacings on the order of 100 m within a simulated tornadic supercell. Variational approaches with and without a vertical vorticity equation constraint are tested, along with a typical (traditional) method involving vertical integration of the mass conservation equation. The analyses employ emulated radar data from dual-Doppler placements 15, 30, and 45 km east of the mesocyclone, with volume scan intervals ranging from 10 to 150 s. The effect of near-surface data loss is examined by denying observations below 1 km in some of the analyses. At the longer radar ranges and when no data denial is imposed, the “traditional” method produces results similar to those of the variational method and is much less expensive to implement. However, at close range and/or with data denial, the variational method is much more accurate, confirming results from previous studies. The vorticity constraint shows the potential to improve the variational analysis substantially, reducing errors in the w retrieval by up to 30% for rapid-scan observations (≤30 s) at close range when the local vorticity tendency is estimated using spatially variable advection correction. However, the vorticity constraint also degrades the analysis for longer scan intervals, and the impact diminishes with increased range. Furthermore, analyses using 30-s data also frequently outperform analyses using 10-s data, suggesting a limit to the benefit of increasing the radar scan rate for variational DDA employing the vorticity constraint.

The Storm Prediction Center (SPC) tornado database, generated from NCEI’s Storm Data publication, is indispensable for assessing U.S. tornado risk and investigating tornado–climate connections. Maximizing the value of this database, however, requires accounting for systemically lower reported tornado counts in rural areas owing to a lack of observers. This study uses Bayesian hierarchical modeling to estimate tornado reporting rates and expected tornado counts over the central United States during 1975–2016. Our method addresses a serious solution nonuniqueness issue that may have affected previous studies. The adopted model explains 73% (>90%) of the variance in reported counts at scales of 50 km (>100 km). Population density explains more of the variance in reported tornado counts than other examined geographical covariates, including distance from nearest city, terrain ruggedness index, and road density. The model estimates that approximately 45% of tornadoes within the analysis domain were reported. The estimated tornado reporting rate decreases sharply away from population centers; for example, while >90% of tornadoes that occur within 5 km of a city with population > 100 000 are reported, this rate decreases to

As convection-allowing ensembles are routinely used to forecast the evolution of severe thunderstorms, developing an understanding of storm-scale predictability is critical. Using a full-physics numerical weather prediction (NWP) framework, the sensitivity of ensemble forecasts of supercells to initial condition (IC) uncertainty is investigated using a perfect model assumption. Three cases are used from the real-time NSSL Experimental Warn-on-Forecast System for Ensembles (NEWS-e) from the 2016 NOAA Hazardous Weather Testbed Spring Forecasting Experiment. The forecast sensitivity to IC uncertainty is assessed by repeating the simulations with the initial ensemble perturbations reduced to 50% and 25% of their original magnitudes. The object-oriented analysis focuses on significant supercell features, including the mid- and low-level mesocyclone, and rainfall. For a comprehensive analysis, supercell location and amplitude predictability of the aforementioned features are evaluated separately. For all examined features and cases, forecast spread is greatly reduced by halving the IC spread. By reducing the IC spread from 50% to 25% of the original magnitude, forecast spread is still substantially reduced in two of the three cases. The practical predictability limit (PPL), or the lead time beyond which the forecast spread exceeds some prechosen threshold, is case and feature dependent. Comparing to past studies reveals that practical predictability of supercells is substantially improved by initializing once storms are well established in the ensemble analysis.

A goal of Warn-on-Forecast (WoF) is to develop forecasting systems that produce accurate analyses and forecasts of severe weather to be utilized in operational warning settings. Recent WoF-related studies have indicated the need to alleviate storm displacement errors in both analyses and forecasts. A potential solution to reduce these errors is the feature alignment technique (FAT), which mitigates displacement errors between observations and model fields while satisfying constraints. This study merges the FAT with a local ensemble transform Kalman filter (LETKF) and uses observing system simulation experiments (OSSEs) to vet the FAT as a potential alleviator of forecast errors arising from storm displacement errors. An idealized truth run of a supercell on a 250-m grid is used to generate pseudoradar observations, which are assimilated onto a 2-km grid using a 50-member ensemble to produce analyses and forecasts of the supercell. The FAT uses composite reflectivity to generate a 2D field of displacement vectors that is used to align the model variables with the observations prior to each analysis cycle. The FAT is tested by displacing the initial model background fields from the observations or modifying the environmental wind profile to create a storm motion bias in the forecast cycles. The FAT–LETKF performance is evaluated and compared to that of the LETKF alone. The FAT substantially reduces errors in storm intensity, location, and structure during data assimilation and subsequent forecasts. These supercell OSSEs provide the foundation for future experiments with real data and more complex events.

Observational and model resolution limitations currently preclude analysis of the smallest scales important to numerical prediction of convective storms. These missing scales can be recovered if the forecast model is integrated on a sufficiently fine grid, but not before errors are introduced that subsequently grow in scale and magnitude. This study is the first to systematically evaluate the impact of these initial-condition (IC) resolution errors on high-resolution forecasts of organized convection. This is done by comparing high-resolution supercell simulations generated using identical model settings but successively coarsened ICs. Consistent with the Warn-on-Forecast paradigm, the simulations are initialized with ongoing storms and integrated for 2 h. Both idealized and full-physics experiments are performed in order to examine how more realistic model settings modulate the error evolution. In all experiments, scales removed from the IC (wavelengths < 2, 4, 8, or 16 km) regenerate within 10–20 min of model integration. While the forecast errors arising from the initial absence of these scales become quantitatively large in many instances, the qualitative storm evolution is relatively insensitive to the IC resolution. It therefore appears that adopting much finer forecast (e.g., 250 m) than analysis (e.g., 3 km) grids for data assimilation and prediction would improve supercell forecasts given limited computational resources. This motivates continued development of mixed-resolution systems. The relative insensitivity to IC resolution further suggests that convective forecasting can be more readily advanced by improving model physics and numerics and expanding extrastorm observational coverage than by increasing intrastorm observational density.

An advection correction procedure is used to mitigate temporal interpolation errors in trajectory analyses constructed from gridded (in space and time) velocity data. The procedure is based on a technique introduced by Gal-Chen to reduce radar data analysis errors arising for the nonsimultaneity of the data collection. Experiments are conducted using data from a high-resolution Cloud Model 1 (CM1) numerical model simulation of a supercell storm initialized within an environment representative of the 24 May 2011 El Reno, Oklahoma, tornadic supercell storm. Trajectory analyses using advection correction are compared to traditional trajectory analyses using linear time interpolation. Backward trajectories are integrated over a 5-min period for a range of data input time intervals and for velocity-pattern-translation estimates obtained from different analysis subdomain sizes (box widths) and first-guess options. The use of advection correction reduces trajectory end-point position errors for a large majority of the trajectories in the analysis domain, with substantial improvements for trajectories launched in the vicinity of the model storm’s gust front and in bands within the rear-flank downdraft. However, the pattern-translation components retrieved by this procedure may be nonunique if the data input time intervals are too large.

The forcing and origins of an internal rear-flank downdraft (RFD) momentum surge observed by the second Verification of the Origin of Rotation in Tornadoes Experiment (VORTEX2) within a supercell occurring near Dumas, Texas, on 18 May 2010 is assessed through ensemble Kalman filter (EnKF) storm-scale analyses. EnKF analyses are produced every 2 min from mobile Doppler velocity data collected by the Doppler on Wheels and Shared Mobile Atmospheric Research and Teaching radars, as well as radial velocity and reflectivity data from the KAMA (Amarillo, Texas) WSR-88D. EnKF analyses are found to reproduce the structure and evolution of an internal RFD momentum surge observed in independent mobile Doppler radar observations. Pressure retrievals of EnKF analyses reveal that the low-level RFD outflow structure is primarily determined through nonlinear dynamic perturbation pressure gradient forcing. Horizontal acceleration into a trough of low perturbation pressure between the low-level mesocyclone and mesoanticyclone and trailing the primary RFD gust front is followed by an abrupt deceleration of air parcels crossing the trough axis. This deceleration and associated strong convergence downstream of the pressure trough and horizontal velocity maximum are indicative of an internal RFD momentum surge. Backward trajectory analyses reveal that air parcels within the RFD surge originate from two source regions: near the surface to the north of the low-level mesocyclone, and in the ambient flow outside of the storm environment at a height of approximately 2 km.

The Warn-on-Forecast (WoF) program aims to deploy real-time, convection-allowing, ensemble data assimilation and prediction systems to improve short-term forecasts of tornadoes, flooding, lightning, damaging wind, and large hail. Until convection-resolving (horizontal grid spacing Δx < 100 m) systems become available, however, resolution errors will limit the accuracy of ensemble model output. Improved understanding of grid spacing dependence of simulated convection is therefore needed to properly calibrate and interpret ensemble output, and to optimize trade-offs between model resolution and other computationally constrained parameters like ensemble size and forecast lead time. Toward this end, the authors examine grid spacing sensitivities of simulated supercells over Δx of 333 m–4 km. Storm environment and physics parameterization are varied among the simulations. The results suggest that 4-km grid spacing is too coarse to reliably simulate supercells, occasionally leading to premature storm demise, whereas 3-km simulations more often capture operationally important features, including low-level rotation tracks. Further decreasing Δx to 1 km enables useful forecasts of rapid changes in low-level rotation intensity, though significant errors remain (e.g., in timing). Grid spacing dependencies vary substantially among the experiments, suggesting that accurate calibration of ensemble output requires better understanding of how storm characteristics, environment, and parameterization schemes modulate grid spacing sensitivity. Much of the sensitivity arises from poorly resolving small-scale processes that impact larger (well resolved) scales. Repeating some of the 333-m simulations with coarsened initial conditions reveals that supercell forecasts can substantially benefit from reduced grid spacing even when limited observational density precludes finescale initialization.