Tornado damage prediction using machine learning

Tornadoes are the most violent of all atmospheric windstorms. Although they occur throughout the world, the highest frequency is observed in the US, where over 1,700 tornadoes were recorded in 2024 alone. Intensity measures that characterize tornado damage, such as the Enhanced Fujita (EF) scale, are derived from post-event assessment on the ground. This somewhat circular aspect of damage characterization and EF scale determination remains a major limitation for long-term prediction of damage. Recognizing the inherent limitations of the standard EF scale, an engineering-relevant hazard characterization is proposed based on a data-driven machine learning (ML) framework to predict the probability that a tornado’s maximum damage wind speed (DWS) will exceed specific thresholds within the EF3 range (136–165 mph). It quality controls and cross-validates records from three primary federal datasets: the Storm Events Database (SED), the Damage Assessment Toolkit (DAT), and the Storm Prediction Center (SPC) database, yielding a curated historical dataset of over 630 authenticated EF3–EF5 tornado events in the United States from 2007 to 2025. Using a Random Forest (RF) classification framework, this study predicts tornado DWS exceedance from four geometric features of the tornado path—track length, maximum width, and the sine and cosine components of path bearing—and evaluates model generalization using State-specific tornadoes held out from model development. The predictive methodology is centered heavily on the data from the State of Iowa (IA) and subsequently tested across various geographical States and regions to examine spatial transferability. The results identify path width as the most influential predictor of severe tornado intensity among the predictors examined, while differences in predictive performance among States indicate that State-specific models may better capture States variability than a universal cross-State model.

In 2024, the United States experienced 27 separate billion-dollar weather and climate disasters, with tornadoes and thunderstorms accounting for more than half of these events (NCEI, 2025). In meteorology, tornado intensity has traditionally been inferred from post-event damage surveys and Enhanced Fujita (EF) scale ratings, with more recent efforts incorporating damage path characteristics and Doppler radar observations (e.g., (Smith et al., 2020a; Smith et al., 2020b)) to support real-time warning operations. These approaches focus on near-real-time forecasts rather than long-term probabilistic characterization. Conversely, wind engineers require long-term predictions of intensity based upon peak parameter values for the built and natural environment design standards. Notably, the majority of tornado-induced billion-dollar losses are associated with the Enhanced Fujita (EF)-3 scale and above tornadoes, denoted here as EF3+, which produce extreme wind speeds and cause significant structural damage. Although these high-intensity tornadoes represent a relatively small fraction of all tornado occurrences, they dominate economic losses, infrastructure disruption, and risk to life safety.

(Twisdale and Dunn, 1983) developed a methodology for assessing tornado risk, which was useful but relied on traditional statistics. (Brooks, 2003) examined F intensity data from 1955 to 2001 to assess the influence of tornado length and maximum width on intensity. He found that the Weibull distribution was appropriate for the tornado length and maximum width, respectively, for each set of the F scale values. (Brooks et al., 2003) designated the occurrence of a tornado per day at a location as a “tornado day”. They identified the mean “tornado day” per year as well as the annual probability of a tornado day occurring anywhere in the US using the 1980 to 1999 NWS database. (Elsner et al., 2014) expanded this work using data for storms from 2007 to 2013. The width and length of tornadoes were fitted with statistical models describing the intensity with wind speed ranges associated with each EF scale. They were able to predict wind speed values for several tornadoes using the derived model with the ultimate goal of identifying an index of tornado destructiveness for all data sets. This index was said to be an improvement of the EF intensity scale. However, the study did not provide improved estimates of the maximum damage wind speed (DWS) or evaluation of the associated pressure drop data required for structural engineering analysis. In response to these limitations, data-driven models have been explored recently.

In particular, machine learning (ML) methods have been applied to wind engineering-related tornado research, e.g., (Jacquot, 2017; Haggag et al., 2022; Zhang et al., 2025). (Jacquot, 2017) used the random forest (RF) method for data from the Storm Event Database SED (NCEI, 2026b) from 2007 to 2017 for the entire US to predict the EF scale of a tornado based on over a dozen characteristics such as location, time, date, etc. The imbalance of the dataset with more EF1 events than EF5 ones was noted as an issue with fitting. He found that the best predictors of tornado damage are width, total costs, and the longitude and latitude of the beginning point. Despite the encouraging results, it was not clear if the data had been cleaned appropriately for intensity scale, width, State, and other parameters. However, the analysis was an excellent start to tornado damage analysis combining quantitative and qualitative data.

(Haggag et al., 2022) used ML for deriving a data-driven model for climate-induced damage prediction as a means to assist community planners. The events included multiple hazards in addition to tornadoes. In addition to the weather data used here, land cover, daily minimum and maximum temperatures, social, housing, demographic and economic attributes were employed. Filter methods were required due to the large number of (sometimes correlated) input features. Several ML and ensemble techniques were used, including regression trees for continuous outputs and classification trees for categorical outputs. Artificial neural networks (ANN) and RF techniques were employed. Because the wind speeds in this investigation were for non-tornadic winds, higher wind speeds regardless of duration caused more property damage.

(Zhang et al., 2025) specifically targeted tornadoes in their investigation. Their goal was to employ data representing hazard, exposure and vulnerability. In particular, the Structural Extreme Event Reconnaissance (StEER) 1 data for fifteen storms were used. The StEER damage degree metric is based on descriptors such as “no visible exterior damage” and “total loss”. Characteristics of building features were also used. The GraphSAGE 2 inductive framework developed at Stanford University was employed due to its efficient scalable learning on large evolving graphs. Unfortunately, all of the performance metrics for accuracy and precision were between 40% to 54%. These scores are considered poor for general ML applications. These studies made it clear that data sorting and feature examination were a crucial first step in ML analysis.

The objective of the present study is to develop and evaluate a data-driven machine learning framework for estimating tornado DWS using post-event tornado damage records. Specifically, this research investigates the application of the RF algorithm to identify the relationships between tornado path characteristics and the probability that a tornado exceeds selected DWS thresholds. The proposed framework estimates exceedance probabilities for specific DWS, providing a continuous representation of tornado intensity within individual EF-scale categories. The framework is developed using authenticated tornado records collected after the implementation of the Enhanced Fujita scale in 2007 and is evaluated through independent validation. In addition to assessing predictive performance, this study examines the relative importance of tornado path characteristics in different States and evaluates the transferability of State-specific models. The ultimate goal is to provide a probabilistic framework that supports engineering applications, including tornado hazard characterization, fragility analysis, and future development of regional tornado risk assessment methodologies.

The paper is organized as follows: Section 2 Materials and Methods presents background on the datasets and their sources of uncertainty, as well as verification with earlier studies. Insights into the tornado parameters of most interest to the wind engineering community are identified and examined. This will be followed by an introduction to the ML methods used for the investigation. Section 3 Results presents the findings of the ML modeling for individual States such as Iowa (IA) as well as nearby multi-State regions. In Section 4 Discussion, the application of the results to structural wind engineering analysis is shown. In Section 5 Summary recommendations and conclusions are provided.

The primary objective of this study was to fit a model for EF3+ tornadoes to predict the likelihood of the maximum damage wind speed exceeding certain values. That is, the maximum damage wind speed represents a more finely tuned metric with a range of wind speeds than the EF scale, e.g., 136–165 mph per EF3, than only fitting the five EF categories. Damage wind speed is a more meaningful metric for wind engineers because measurements of site-specific surface level wind speeds are not available and the vulnerability of structural components is captured better with a pressure loading derived from a damage wind speed versus a broad category (EF) that encompasses multiple wind speeds. Estimation of the damage wind speed requires a more precise analysis of the post-event location site than EF scale. The EF-related degree of damage scale typically leads to a damage wind speed prediction based on engineering analysis. This estimation may provide data regarding the vulnerability of the structural components whereas the designation of “EF3” for an event does not.

First, in this paper, the data were identified and cleaned for use in the investigation. Second, various types and numbers of features, such as tornado length and width, were used to assess the best set for modeling. Thirdly, the RF classification and logistic regression calibration models were explored and selected for this study. Fourth, developments and improvements were made, including developing State-specific RF models, balancing the dataset to address data imbalance, and independent validation to show the RF model is working on that State tornado data. In this section, an overview of the data and methods are described.

Datasets used

Figure 1 illustrates a tornado funnel model with a low-rise building model juxtaposed with a typical structural engineering wind loading diagram. The tornado model shows the maximum tangential velocity and the translational speed. Although the funnel is drawn as narrower at the ground level, suggesting smaller wind speeds, it has been shown that the surface level tornado wind speeds are greatest near the ground level, which is the opposite of boundary layer flow diagrams for straight line winds (Kosiba and Wurman, 2023). The maximum DWS is not the vector sum of the measured tornado wind speeds, but rather the estimated wind speed assumed to cause the damage to the built and natural environment according to the EF scale e.g. (Wind Science and Engineering Center WISE, 2004). Unfortunately, measuring the surface wind speed during a tornado event is extremely challenging and relating the radar-recorded values of extreme storms to surface level winds has not been completely successful to date. Further, identifying the associated atmospheric pressure drop values and impact on low-rise buildings along the tornado path has not been successful primarily due to recording challenges.

It is noted that categorial EF scales display uncertainty including but not limited to post-event observer bias; use of building standards and construction methods at the site that vary from State to State; lack of information about the atmospheric pressure drops at the site; and the age of the built environment. However, it is assumed that the collection of post-event data by observers in each State has resulted in more consistency at the State level versus the US as a whole. Because a range of associated DWS are provided for each EF scale, the assumption made here is that the DWS estimate (the 3-s gust) is a better representation of the damage than the categorical levels for EF 1 through 5 as mentioned previously. US post-event records kept from 2007 provide this data, which has been authenticated. Even so, the dozen or more parameters possibly influencing the damage make long term predictions complex. Therefore, the use of modeling approaches amenable to complex categorical and numerical data was examined, resulting in the decision to explore ML.

Post-event datasets exist for tornado damage analysis, but making long term predictions is complex. Over a dozen parameters are used to characterize events, and damage is assumed to be correlated with building code standards by State. That is, some buildings are constructed for high winds in States subjected to hurricanes, which may overlap with tornado prone areas. It is assumed these buildings would resist higher tornado speeds than for other regions without strong wind-based codes. This assumption was considered in the sorting of the datasets.

Post-event, in situ tornado data collected by the National Weather Service (NWS), including local weather service event reports and supporting news media reports, were used in this investigation. The field-collected datasets are derived primarily from first-hand post-event damage surveys and provide the basis for the ML analyses presented in this study.

Three primary tornado datasets were employed: (1) the Storm Events Database (SED) maintained by the National Centers for Environmental Information NCEI (NCEI, 2026b) (2) the Damage Assessment Toolkit (DAT) maintained by (NCEI, 2026a); and (3) the Storm Prediction Center (SPC) tornado database maintained by the National Weather Service NWS (NWS, 2024). The SED dataset is available in CSV format through NCEI. SPC tornado records in CSV format are provided by the NOAA/NWS Storm Prediction Center in Norman, Oklahoma. The DAT dataset was accessed through NCEI Geographic Information System (GIS) services using the Damage Assessment Toolkit interface. Although the three datasets overlap and complement one another, the DAT database provides the most comprehensive post-event damage information and therefore served as the primary data source for this investigation.

The DAT dataset contains three categories of information: damage path data, damage point data, and damage polygon data. Damage path data summarize the overall characteristics of each tornado, including location, occurrence time, start and end coordinates, path length, maximum path width, fatalities, injuries, EF rating, estimated maximum DWS, property damage, crop damage, and the responsible NWS Weather Forecast Office. Damage point data provide detailed observations at individual damage locations, including the Damage Indicator, Degree of Damage, estimated DWS, and a brief description of the observed damage. Damage polygon data provide estimated spatial extents of tornado damage. Because only a limited number of polygon records were available and the mapped damage areas were approximate, the polygon data were not used in the present study.

Additional geometric parameters were derived from the DAT damage path data to support the ML models. These include tornado bearing angle, tornado duration, and the straight-line distance between the start and end locations. In this study, the bearing angle is defined as the direction of tornado movement measured clockwise from geographic north. Because bearing angle is a circular variable, its sine and cosine components were subsequently used as predictor variables in the RF models.

The geometric characteristics of tornado damage paths were also examined. Tornado path width and intensity vary along the damage path, although only the maximum path width is routinely recorded in the available datasets. Figure 2 illustrates the geometric variables used in this investigation and defines the damage-path coordinate system adopted for the analyses. The variable x D represents the damage distance along a line connecting the beginning and end points; y D represents the normal of the x D line to the damage point recorded in the dataset. In this manner, the deviation from the beginning to end can be plotted to examine tornado behavior beyond recording the maximum DWS. Two (x D, y D, DWS) plots for EF4 tornadoes are provided for Iowa and Mississippi, respectively in Figure 2. The y D axis indicates whether the tornado moved left (positive) or right (negative) along its path from beginning to end. It can be seen that the tornado properties differ for the two states as first noted by Fujita (1978). That is, tornado damage paths vary from event to event, including straight, curved, left-turning, right-turning, meandering, and U- or V-shaped paths. Investigations of the behaviors shown in Figure 2 for EF4 tornadoes are beyond the scope of this paper and will be examined in future studies. The maximum DWS and maximum width were used in this study.

Previous studies have also shown that tornado intensity varies considerably along the damage path (e.g., Standohar-Alfano and van de Lindt, 2015). Tornadoes frequently intensify from lower EF ratings to peak intensity before weakening toward dissipation. Although graphical representations based on changing EF-scale distributions along the damage path have been proposed in previous studies, the tornadoes examined in the present dataset did not exhibit sufficiently consistent spatial patterns for those methods to be adopted. However, one observation from the preliminary analyses was that the translational speed of many tornadoes was approximately one mile per minute (26.8 m/s) across the States examined.

Before model development, the three datasets were merged and quality controlled. The tornado records were divided into two groups: EF3 and above tornadoes and EF2 and below tornadoes, which are denoted as EF3+ and EF2-, respectively. Different data-processing procedures were applied to these two groups because the maximum DWS were required for EF3+ tornadoes but was not required for EF2- tornadoes in this study. It is important to clarify that this was not designed to filter out tornado records based on predefined quality criteria. The primary objective was to retain all identified tornado events while ensuring that each record represents a single, complete, and correctly characterized tornado event in the United States. Records with missing, inconsistent, duplicated, or potentially erroneous information were therefore investigated and cross-referenced among the available data sources. Where sufficient information was available, these records were corrected, supplemented, or consolidated rather than simply removed.

The primary source for EF3+tornadoes was the DAT because it provides a corresponding maximum DWS for individual tornadoes. DAT tornado records from 2007 through 2025 were downloaded and subjected to quality control. The variables examined included tornado length, width, EF scale, maximum DWS, and duration. Records containing missing, invalid, or inconsistent values in these fields were flagged for further investigation rather than immediately removed. Particular attention was given to tornadoes crossing State boundaries, where differences in reporting practices and time zones occasionally produced discrepancies in event duration and location. These records were retained because the corresponding tornadoes were known to have occurred, and the identified errors could result from incorrect labeling or reporting in the original records.

Because DAT does not contain a complete record of all EF3+ tornadoes that occurred in the United States between 2007 and 2025, SPC tornado records were used to identify missing events and cross-check the DAT records. SPC data containing U.S. tornado records from 2007 through 2025 were downloaded and quality controlled by removing duplicate records and checking for missing, invalid, or inconsistent values. Compared with DAT, the SPC dataset is more systematically organized, and incorrectly labeled records were relatively uncommon. The duplicate records in SPC are a result of their unique recording style: when a tornado crosses several counties or States, SPC tends to record them as several separated tornadoes. The number of segments depends on the number of counties and/or States this tornado crosses. The record of the entire tornado is also provided as well. In this study, only the record of the entire tornado is required, the records of the segments need to be removed. The cleaned SPC and DAT datasets were then cross-referenced using tornado occurrence date, location, and other available identifying information. This cross-checking served two purposes: (1) correcting mis-labeled DAT records and (2) identifying and adding EF3+ tornadoes that were missing from DAT. When discrepancies were identified, additional credible sources, including available field survey reports and event summaries from the NWS and NOAA, were consulted to further cross-validate the records. Modifications were made only when the information was supported by at least two independent sources.

Because SPC does not provide the maximum DWS associated with each tornado, SED was subsequently used to obtain additional information. The SPC and SED tornado records are derived from the same underlying storm-event records and contain the same tornado information. However, SED additionally provides event and episode narratives for individual tornadoes, whereas these narratives are not included in the SPC bulk-download dataset. These narratives were reviewed to identify reported maximum DWS when available. In addition to the SED, field survey reports and event summaries published by the NWS and NOAA are also used to identify the maximum DWS.

For tornadoes for which the maximum DWS could not be identified from the SED narratives, field survey reports and event summaries published by the NWS and NOAA, the maximum DWS was assigned conservatively based on the tornado’s EF scale. Specifically, an EF3 tornado was assigned a maximum DWS of at least 136 mph, and an EF4 tornado was assigned a maximum DWS of at least 165 mph. This approach was used for cases in which historical records were incomplete or a specific maximum DWS was unavailable, including some older tornadoes and tornadoes for which maximum DWS were not reported, such as tornadoes in Montana State.

The EF2- tornado dataset was based primarily on SPC records because an individual maximum DWS was not required for these tornadoes in this study. SPC tornado records from 2015 through 2025 were downloaded and subjected to the same basic quality-control procedures, including removal of duplicate records and checking of missing, invalid, or inconsistent values. When an anomalous or potentially incorrectly labeled record was identified, the record was cross-checked against other credible sources. Corrections were only made when two or more other credible sources show support evident that the original SPC record was incorrect.

Following these quality-control and cross-validation procedures, the final ML dataset consisted of authenticated tornado damage-path records from 2007 (first use of EF scale) through 2025. The final database contains more than 600 authenticated EF3–EF5 tornadoes across the United States together with the corresponding lower-intensity tornadoes used for State-specific model development.

Random forest model framework for tornado damage wind speed

Model selection and development

The RF classifier was selected as the primary modeling approach in this study based on previous studies by Jacquot (2017) and Zhang et al. (2025), as well as the suitability of RF models for the application. RF models are ensemble learning methods that construct multiple decision trees and combine their predictions to improve classification performance and reduce overfitting (Breiman, 2001). Each decision tree recursively partitions the data by selecting predictor variables that best separate the target classes at each node, and the final prediction is obtained by aggregating the predictions from all trees in the forest. In this study, RF models were used to predict whether a tornado’s maximum DWS exceeds selected thresholds based on tornado path geometry characteristics. To provide baseline comparisons with the proposed RF model, Logistic Regression (LR) and Extreme Gradient Boosting (XGBoost) models were also developed. LR is a linear classification method that estimates the probability of a binary outcome using the logistic function (Hosmer et al., 2013), whereas XGBoost is an ensemble learning method that builds an additive sequence of decision trees using gradient boosting (Chen and Guestrin, 2016).

For model development, the available dataset was randomly divided into training and testing subsets, with 80% of the data used for training and 20% reserved for testing. The testing subset was held out from model fitting and hyperparameter optimization and was used only to evaluate model performance. The same training and testing datasets were used for the RF, LR, and XGBoost models to ensure a consistent basis for comparison. To optimize model performance, hyperparameter tuning (Bergstra and Bengio, 2012) was performed on all three types of models using the training dataset.

Feature selection

The following features were selected for use in the ML model: tornado path length, tornado path width, and the sine and cosine values of the tornado bearing angle. Several considerations motivated the selection of only these four variables. First, previous studies by (Jacquot, 2017; Zhang et al., 2025) showed that many variables available in tornado datasets contribute little to model prediction. Although those studies included a larger number of features, feature importance analyses indicated that only a small subset had substantial influence on model performance. Second, this study aimed to develop a practical and transferable framework. Therefore, the number of model features was intentionally minimized while maintaining strong predictive performance. Third, several available tornado parameters are highly correlated with one another. To reduce redundancy and multicollinearity, correlated variables were excluded from the model. For example, tornado duration is strongly correlated with tornado path length, and tornado path area is also closely related to path length and width. Consequently, these variables were excluded. Fourth, some variables were excluded because of the modeling framework itself. Temporal information, including tornado year, month, and day, was excluded because separate models were developed for different time periods. Similarly, geographic location variables, such as State identifiers, were excluded because separate State-specific models were constructed. Variables related to fatalities, injuries, property damage, and crop damage were also excluded, as a considerable portion of tornado records lacks complete fatality and injury information. In addition, property and crop damage values are reported in nominal dollar amounts corresponding to different years. Because of inflation and currency value changes over time, these values are not directly comparable without additional normalization. Finally, the tornado bearing angle was represented using its sine and cosine values rather than the angle itself. Angular measurements in degrees introduce discontinuities because 0° and 360° represent the same direction, which can negatively affect machine learning model performance. Although the bearing angle in radians was also tested, the sine and cosine representation produced better predictive results. Therefore, the sine and cosine values of the bearing angle were selected as the final directional features. Through preliminary testing, it was determined that tornado path length, tornado path width, and the sine and cosine values of the tornado bearing angle form the basic geometry and the best representation of the tornado damage path.

Damage wind speed threshold selection

In this study, the maximum DWS was predicted rather than EF-scale classifications or building damage levels, as discussed previously. Each EF-scale category represents a range of DWS rather than a single value. For example, an EF3 tornado corresponds to maximum DWS between 136 (60.8 m/s) and 165 mph (73.8 m/s). However, tornadoes with maximum DWS near 136 mph can produce substantially different levels of damage compared to tornadoes with maximum DWS near 165 mph. Therefore, directly predicting EF-scale categories would ignore important variations in damage severity within the same EF-scale classification. To better capture these variations, this study predicts exceedance of selected DWS thresholds.

An alternative approach would be to directly predict the Degree of Damage 3 to structures. Such an approach could potentially avoid uncertainties associated with EF-scale and DWS assignments, which are determined by post-event damage surveys. However, predicting the Degree of Damage introduces additional challenges because structural damage depends on many factors beyond the tornado itself. These factors include building type and size, structural quality, construction materials, building orientation relative to the tornado path, and the location of the structure within the tornado damage field. In addition, building codes vary significantly among States and regions. For example, structures in IA must account for snow loads, whereas buildings in MS are designed for different weather considerations. As a result, models trained using building damage data from one State may not generalize well to other States with different construction practices and environmental conditions.

The selected DWS thresholds in this study are 136, 140, 145, 150, 155, 160, and 165 mph 4. These thresholds are all within the EF3 intensity range. The primary reason for selecting these thresholds is that they correspond to the range of tornado damage levels of greatest interest in this study. Under current building codes, most structures are generally expected to withstand EF2-and-below tornadoes, which correspond to maximum DWS of 135 mph or lower. In contrast, EF4 and EF5 tornadoes, with maximum DWS exceeding 166 mph, often produce near-total destruction, typically corresponding to Degree of Damage of eight and above. Therefore, the 136–165 mph range represents an important transition zone in which tornadoes can produce a broad spectrum of severe but not necessarily complete structural damage.

In addition, EF3 tornadoes occur more frequently than EF4 and EF5 tornadoes. As a result, tornadoes with maximum DWS between 136 and 165 mph are more likely to be encountered in real-world hazard scenarios than tornadoes exceeding 166 mph. The greater occurrence frequency of EF3 tornadoes also provides a substantially larger dataset for analysis and model development. This improves the statistical robustness of the study and allows the proposed framework to be trained and validated using a sufficient number of tornado events.

Dataset structure and selection

Dataset imbalance is one of the major challenges in tornado damage prediction. Tornado data are highly imbalanced because violent tornadoes are relatively rare compared to weaker tornadoes. For example, only 10 EF5 tornadoes occurred in the United States between 2007 and 2025, while more than 10,000 tornadoes were recorded during the same period. Tornadoes rated EF3 and above account for less than 8% of all recorded tornadoes in the United States. This imbalance can cause ML models to overpredict high intensity tornadoes (EF3 to EF5), leading to excessive false positive predictions. To address this issue, the dataset structure was adjusted to increase the representation of tornadoes above the selected thresholds while limiting the number of tornadoes below the thresholds.

Therefore, all tornadoes in the United States with maximum DWS exceeding the selected thresholds between 2007 and 2025 are included in the above-threshold dataset.

IA State was selected for the first in-depth investigation as it is the location of the National Testing Facility for Enhancing Wind Resiliency of Infrastructure in Tornado-Downburst-Gust Front Events (NEWRITE) project. The NEWRITE team is designing and ultimately constructing a mid-scale experimental facility for research, education and industrial testing of mid-to-large scale structures in tornado, downburst and gust-front like winds at Iowa State University. IA State is part of the midwestern tornado alley. For the IA model, tornadoes below the thresholds from 2015 to 2025 were selected. Two primary reasons motivated this choice.

The first reason is dataset balance and model performance. Approximately 600 tornadoes across the United States exceeded the 136 mph threshold during the study period. To maintain a relatively balanced dataset, a similar number of below-threshold tornadoes is desired. Between 2015 and 2025, IA recorded approximately 700 tornadoes, including roughly 680 EF2-and-below tornadoes and approximately 20 EF3-and-above tornadoes. As a result, selecting IA tornadoes from 2015 to 2025 provides a relatively balanced dataset for the 136 mph threshold. As the threshold increases, the number of above-threshold tornadoes decreases, resulting in greater imbalance. However, for consistency across all threshold analyses, the below-threshold IA tornado dataset remains fixed between 2015 and 2025. In these cases, Synthetic Minority Over-sampling Technique SMOTE (Chawla et al., 2002) was additionally applied to mitigate imbalance effects when necessary.

The second reason for selecting the 2015–2025 period is data quality and reliability. Damage reports and event summaries from sources such as the National Weather Service (NWS), National Oceanic and Atmospheric Administration (NOAA), and National Centers for Environmental Information (NCEI) are generally available for most EF3+ tornadoes between 2007 and 2025. These records allow cross-validation among multiple data sources: DAT, SED, and SPC, when discrepancies occur. In contrast, weaker tornadoes such as EFU, EF0, and EF1 events often lack detailed damage reports and event summaries, making cross-validation more difficult. Earlier records, particularly from 2007 to 2010, are more likely to contain incomplete or missing information. Therefore, selecting IA tornadoes from 2015 to 2025 improved overall dataset reliability and reduced uncertainty associated with incomplete historical records.

Independent validation

An independent validation dataset was introduced to evaluate the generalization capability of the proposed framework. Because the training and testing datasets are intentionally adjusted to improve class balance by introducing above-threshold tornadoes from outside the target State, independent validation is necessary to assess whether the RF model can accurately classify both above-threshold and below-threshold tornadoes within the target State.

For the IA RF models, the validation dataset consisted of both above-threshold and below-threshold IA tornadoes that were not directly used during model training or testing. Specifically, all IA tornadoes exceeding the selected threshold between 2015 and 2025 are excluded from the training and testing datasets. In addition, five times as many below-threshold IA tornadoes from the same period are randomly selected and also excluded from the RF model training and testing. Together, these tornadoes form the independent validation dataset. The same validation strategy is applied consistently to all State RF models developed in this study. This independent evaluation provides a more comprehensive assessment of the RF model’s predictive capability. Because the validation tornadoes are fully separated from the training and testing processes, the validation results provide a more reliable estimate of the RF model’s predictive performance.

Logistic regression calibration

Although the RF classifier produces probability estimates for tornado DWS exceedance, these probabilities are not directly representative of real-world exceedance likelihoods because the training dataset is intentionally rebalanced. Specifically, the dataset is enriched with above-threshold tornadoes in order to improve model learning and reduce the effects of severe class imbalance. As a result, the proportion of above-threshold tornadoes in the training dataset is substantially higher than the true occurrence rate observed in real tornado populations. Because the model is trained in this above-threshold-enriched environment, the raw probability outputs from the RF classifier may systematically overestimate the real-world probability of exceedance. Therefore, probability calibration is necessary before the model outputs can be interpreted as realistic exceedance probabilities.

Two calibration approaches are considered in this study. The first approach is the traditional Bayes prior correction method, which adjusts the predicted probabilities based on the difference between the class proportions in the training dataset and the true class proportions in the target population. This method assumes that the conditional distributions of the predictor variables remain consistent between the balanced training dataset and the real-world dataset. In this study, the independent validation dataset provides empirical justification for evaluating whether this assumption is reasonable.

The second approach is logistic regression calibration, following the method proposed by (Dankowski and Ziegler, 2016). Logistic regression is a classification method that estimates the probability of an observation belonging to a particular class based on a set of predictor variables. In this approach, a logistic regression model is fitted using the RF probability outputs as predictors, allowing the calibration process to account for systematic nonlinearities and deviations in the probability estimates without relying strictly on prior probability assumptions.

The IA tornado dataset from 2015 to 2025 was used to fit the calibration model. IA tornado records provide population-specific information for adjusting RF probabilities because IA is the target population. The empirical exceedance rate was calculated as the proportion of IA tornadoes exceeding each specified DWS threshold. The calibrated probabilities from both approaches were compared with these empirical exceedance rates. Although the Bayes prior correction is theoretically valid under its assumptions, logistic regression calibration produced closer agreement with the observed IA exceedance rates across the evaluated thresholds, shown in the Appendix. This may be because logistic regression calibration empirically learns the relationship between the RF probability estimates and the observed IA outcomes, allowing it to correct systematic deviations in the RF probabilities that may not be fully explained by class-prior differences alone. Therefore, logistic regression calibration was selected as the final calibration method.

Because the calibration dataset overlaps with the RF development data, an independent calibration-validation dataset was constructed from 355 IA EF0 to EF5 tornadoes from 2007 to 2014, including 11 EF3+ tornadoes. For the purpose of this independent validation only, 11 IA EF3+ tornadoes were temporarily excluded from the RF training/testing dataset to ensure that these tornadoes were completely unseen by both the RF and calibration models. These 11 tornadoes were removed only for the independent calibration-validation analysis and were not otherwise excluded from the study datasets. In these 11 EF3+ tornadoes, 11 of them exceeding 136 mph and 140 mph, 8 of them exceeding 145 mph and 150 mph, 7 of them exceeding 155 mph, 6 of them exceeding 160 mph, and 5 of them exceeding 165 mph.

Goodness-of-fit

The goodness-of-fit metrics used to evaluate model performance are defined in this section. RF and baseline Model performance is evaluated using accuracy, precision, recall, and F1 score. Accuracy represents the proportion of correctly classified observations among all observations and is defined in Equation 1. Precision measures the proportion of correctly identified positive cases among all cases predicted as positive and is defined in Equation 2. Recall, also referred to as sensitivity, measures the proportion of actual positive cases that are correctly identified by the model and is defined in Equation 3. The F1 score is the harmonic mean of precision and recall and provides a balanced measure of model performance when both false positives and false negatives are important and is defined in Equation 4. The accuracy scores presented in the following sections represent the overall classification accuracy of the RF models, accounting for correct predictions of both below-threshold and above-threshold tornadoes. For binary classification problems, overall accuracy provides a useful measure of the model’s ability to correctly identify tornadoes belonging to both classes. For binary classification, the overall accuracy is equal to the micro precision score for both classes and the micro recall score for both classes according to the equations in the above section.

In addition to accuracy, weighted F1 scores are reported to evaluate overall model performance while accounting for class imbalance within the validation datasets. The weighted F1 score for binary classification models is defined in Equation 5. The weighted F1 score accounts for class imbalance by weighting each class according to its number of observations, providing a more representative overall measure of model performance across both below-threshold and above-threshold tornadoes. Together, these metrics provide a comprehensive assessment of the predictive performance of the RF models.

Class-specific recall scores are also reported for both above-threshold and below-threshold tornadoes. From a hazard assessment and engineering perspective, the ability to correctly identify above-threshold tornadoes is of primary importance because these events are associated with the greatest potential for severe damage. In practice, it is generally preferable for a below-threshold tornado to be incorrectly classified as above-threshold (false positive) than for an above-threshold tornado to be incorrectly classified as below-threshold (false negative). Therefore, recall for the above-threshold class is a particularly important performance metric and is emphasized throughout the analysis.

Because tornadoes from other States were incorporated into the training datasets, independent validation datasets were used to evaluate whether the RF models can successfully identify both above-threshold and below-threshold tornadoes within the target State. For each State model, the independent validation dataset consisted of State-specific tornadoes that were not used during model development. The overall accuracy, weighted F1 score, and class-specific recall scores were reported for both the RF model testing datasets and the independent validation datasets to provide a comprehensive evaluation of model performance and generalization capability.

The results presented in Section 3.2 intend to evaluate whether the developed State RF model can serve as a universal RF model for predicting tornadoes in other States and examine the relationships among tornadoes across different States. Because the number of EF3 and above tornadoes available for cross-validation varies among States and is relatively small for some States, 95% bootstrap confidence intervals (CIs) were calculated for accuracy, class-specific recall, and weighted F1 score using bootstrap resampling to quantify the uncertainty of these performance estimates.

A confusion matrix is used to summarize classification results by comparing the true class labels with the predicted class labels in a matrix format and presented graphically. The confusion matrix provides the numbers of true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN), which form the basis for calculating the evaluation metrics described above. An example of the confusion matrix is shown in the following sections.

Feature importance is a model interpretation technique used to quantify the contribution of each predictor variable to a ML model’s prediction, e.g., (Artificial Intelligence, 2026). In this study, permutation feature importance was used to evaluate the importance of each feature. Permutation feature importance measures the change in model performance when the values of a particular feature are randomly shuffled while all other features remain unchanged. If permuting a feature causes a substantial decrease in model performance, the feature is considered important because the model relies on the information contained in that feature for prediction. Conversely, if permuting a feature results in little change in performance, the feature contributes little predictive information to the model.

Calibration model performance was evaluated using the Brier score, calibration slope, and calibration intercept. The Brier score measures the accuracy of probabilistic predictions, with lower values indicating better overall agreement between predicted probabilities and observed outcomes. The calibration intercept and slope quantify systematic over- or underprediction and the agreement between predicted and observed probabilities, respectively, with an intercept of zero and a slope of one indicating ideal calibration. The calibration slope and intercept are also reported with 95% bootstrap CIs to quantify the uncertainty of the estimates.

Details of hyperparameter optimization, the application of SMOTE, the bootstrap procedure used to estimate CIs for performance metrics, and supplementary model performance metrics are provided in the Supplementary Material.

Table 1 shows the number of data used in each model. Figure 3 shows the regions and States of the US employed for the ML modeling (US Census Bureau, 2025). IA was selected as the initial State for modeling due to its location in tornado alley as well as being the center of the NEWRITE project. The location of the States relative to IA and the Midwest was the primary organizing principle in the order of ML modeling.

Throughout the following sections, a model name in the format “State DWS RF” refers to the RF model developed for a DWS threshold in mph. For example, the IA 136 RF model refers to the RF model used to predict whether a tornado’s maximum DWS is at least 136 mph. Similarly, “State DWS RF Validation” refers to the validation results for that RF model. For example, “IA 136 RF Model IA Validation” refers to validating the IA 136 RF model using the independent IA validation dataset, while “IA 136 RF Model MO Validation” refers to applying the IA 136 RF model to the Missouri validation dataset. The two baseline models follow the naming formats “State DWS LR” and “State DWS XGBoost,” referring to the Logistic Regression and XGBoost models, respectively, developed for the corresponding DWS threshold.

IA model results

Figure 4 shows the confusion matrix for the IA 136 RF model. The matrix shows the larger prediction along the diagonal, which indicates a good fit. Figure 5 summarizes the accuracy, weighted F1 score, and recall scores for both classes for all IA RF models and their corresponding validation datasets. For clarity, the y-axis of some of the bar charts shown starts from 50%. Overall, the IA RF models demonstrate strong predictive performance across all selected DWS thresholds. The accuracy and weighted F1 scores across both the RF models and the independent validation datasets consistently exceed 85%, indicating that the proposed framework can distinguish tornadoes with maximum DWS exceeding the selected thresholds from those that do not.

The recall scores for both above-threshold and below-threshold tornadoes were also generally above 85% for all RF models and validation datasets. These results indicate that the RF models correctly identified a high proportion of tornadoes in both classes. From a hazard assessment perspective, the recall scores for the above-threshold class are particularly important because they quantify the RF model’s ability to identify potentially damaging tornadoes. The consistently high recall values obtained for the above-threshold class demonstrate that the proposed framework can effectively identify tornadoes associated with higher maximum DWS while maintaining strong overall classification performance. Furthermore, the comparable performance observed between the RF model testing datasets and the independent validation datasets suggests that the RF models are learning meaningful relationships between tornado path geometry and maximum DWS rather than simply memorizing the training data.

Figure 6 presents the permutation feature importance values for the IA RF models. Among all predictor variables, tornado path width was consistently the most important feature with substantially greater importance than the remaining variables. This indicates that the predictive performance of the RF models was most sensitive to information contained in tornado path width. When the width variable was randomly permuted, RF model performance decreased more than when any other feature was permuted, demonstrating that width contained the greatest amount of information for distinguishing above-threshold tornadoes from below-threshold tornadoes. It was observed from the existing tornado records that tornadoes with higher maximum DWS generally produce larger and more extensive damage paths, resulting in greater path widths.

Tornado path length was consistently the second most important feature. Although its permutation importance values are considerably smaller than those of path width, they remain substantially greater than those of the bearing-angle variables. This suggests that path length provides meaningful predictive information for distinguishing tornadoes with different maximum DWS. The result is also consistent with existing tornado records, as stronger tornadoes often persist for longer durations and produce longer damage paths. However, the relationship between tornado intensity and path length is not as direct as the relationship with path width. Numerous examples exist in which an EF3 tornado has a shorter path length than an EF2- tornado. Consequently, path length contributes useful predictive information but is less influential than path width.

The sine and cosine values of the tornado bearing angle were generally the third and fourth most important features, although their relative rankings varied among the RF models. In some RF models, the cosine term contributed more strongly, whereas in others the sine term was more influential. The variation in permutation importance among the directional variables suggested that differences exist among the selected DWS thresholds, even though all thresholds fall within the EF3 range. The fact that the RF models assign different levels of importance to the directional variables for different thresholds supported the decision to analyze these DWS thresholds individually rather than treating the entire EF3 range as a single category.

The relatively low permutation importance values associated with the bearing-angle variables were expected for IA tornadoes because the majority of recorded tornadoes in the State exhibit similar southwest-to-northeast movement patterns. As a result, tornado direction provided less discriminatory information than path width or path length when distinguishing between tornadoes with different maximum DWS. Nevertheless, the non-zero importance values indicated that directional characteristics still contribute some predictive information to the RF models.

Figure 5 presents the results of the IA RF models developed using RF, while Figure 7 presents the results of the IA models developed using LR and XGBoost. The results of the IA 136, 140, 145, 150, 155, 160, and 165 models developed using RF, LR, and XGBoost indicate that the three model types produce comparable performance in terms of accuracy and weighted F1 score. Compared with the RF models, the LR models achieve slightly higher recall for the class below the corresponding thresholds, with improvements of approximately 1%–2%. However, this improvement is accompanied by substantially lower recall for the class above the thresholds, with Class 1 recall being at least 5% lower than that of the RF models. In contrast, the XGBoost models generally exhibit performance comparable to the RF models in terms of both overall evaluation metrics and recall for the classes above and below the thresholds. The main exceptions are the IA 145 and IA 165 models, for which the RF models achieve at least 2% higher recall for the class above the threshold than the corresponding XGBoost models. Although the three model types demonstrate broadly comparable performance from a general ML classification perspective, recall for the class above the DWS threshold is particularly important for the wind engineering application of this study, as these cases represent tornadoes with potentially greater hazard risks and associated damages. From this wind application-specific perspective, the RF model provides an advantage by consistently maintaining higher recall for the class above the threshold while achieving comparable overall predictive performance. These results support the selection of RF as the primary modeling approach in this study.

Figure 8 presents the probabilities of exceeding the selected DWS thresholds in IA from 2015 to 2025. Table 2 presents the calibration validation results using the independent IA tornado dataset from 2007 to 2014. The Brier score, calibration slope, and calibration intercept, together with their 95% CIs, were used to evaluate the calibrated RF probabilities. Logistic calibration reduced the Brier score across all evaluated DWS thresholds, indicating improved overall probabilistic performance. The calibration slopes and intercepts indicate that the degree of agreement between predicted and observed probabilities varied among thresholds. These results reflect the performance of the complete RF prediction and calibration framework on a temporally independent dataset and therefore are also influenced by the prediction performance of the underlying RF models on the independent calibration validation dataset. In particular, an EF3+ tornado in the calibration validation dataset was incorrectly classified by the IA 145 RF model, and another EF3+ tornado was incorrectly classified by the IA 150 RF model. In addition, the temporary exclusion of EF3+ tornadoes from the RF training/testing dataset for independent calibration validation further reduced the number of EF3+ tornadoes available for RF model development. These factors may contribute to the threshold-specific variability observed in the independent calibration validation. Thus, while the reduction in Brier scores demonstrates improved overall probabilistic performance after calibration, the calibration slopes and intercepts indicate that some threshold-specific uncertainty remains when the calibrated probabilities are applied to an independent temporal period. These results demonstrate that the proposed modeling framework can be used not only to classify tornadoes according to selected DWS thresholds but also to estimate State-level probabilities of exceedance. Consequently, the framework provides an approach for estimating the likelihood of tornadoes exceeding specified DWS thresholds within a given State.

IA 136 RF model with different states data validation results

Figure 9 presents the validation results of the IA 136 RF model with 95% CIs when applied to tornado datasets from other States. Unlike the previous validation analysis, which used independent IA tornado data, this analysis evaluated whether the IA 136 RF model can serve as a universal RF model for predicting tornadoes in other States. The selected States include those adjacent to IA, States within the Northwest Central Region (Figure 3), and States that experience frequent tornado activity but are geographically distant from IA.

RF Model performance varied substantially across States, indicating that the IA 136 RF model cannot be directly generalized to all States. Illinois, Missouri, Nebraska, Wisconsin, and Minnesota exhibited accuracy scores, weighted F1 scores, and recall scores that were generally comparable to those obtained for IA. For example, Illinois achieved an accuracy of 89% (95% CI: 82%–95%) and a weighted F1 score of 90% (95% CI: 85%–96%), while Missouri achieved an accuracy of 81% (95% CI: 74%–87%) and a weighted F1 score of 83% (95% CI: 78%–89%). These results suggest that these States may share similar relationships between tornado path geometry and maximum DWS. However, the differences in the point estimates and CIs indicate that the strength of this similarity varies among States. Notably, all of these States are geographically adjacent to or near IA.

In contrast, other States exhibited different performance patterns. Kansas, North Dakota, South Dakota, and Florida achieved relatively high overall accuracy and weighted F1 scores but comparatively lower recall for above-threshold tornadoes. For example, the recall for the above-threshold class was 75% (95% CI: 58%–90%) for Kansas, 67% (95% CI: 20%–100%) for North Dakota and South Dakota, and 75% (95% CI: 40%–100%) for Florida. Conversely, Mississippi and Tennessee produced relatively strong recall for above-threshold tornadoes but weaker recall for below-threshold tornadoes, resulting in lower overall accuracy and weighted F1 scores. Texas exhibited weaker performance across both classes compared with IA, with an accuracy of 79% (95% CI: 74%–84%) and a weighted F1 score of 81% (95% CI: 77%–86%). Collectively, these results suggest that the relationship between tornado path geometry and maximum DWS varies among States and that a single universal model may not adequately capture these differences across all States.

The uncertainty associated with these validation results should also be considered when comparing RF model performance among States. The 95% CIs indicate that the precision of the performance estimates varies among States and among performance metrics. States with relatively small numbers of validation events generally exhibited wider CIs, particularly for recall of the above-threshold class. For example, the above-threshold recall for North Dakota and South Dakota was 67%, but the corresponding 95% CI ranged from 20% to 100%, indicating substantial uncertainty in this estimate. Similarly, Florida had an above-threshold recall of 75% with a 95% CI of 40%–100%. In comparison, States with larger validation datasets, such as Mississippi, Tennessee, and Texas, generally exhibited narrower CIs. These results indicate that some of the observed differences in RF model performance among States may be influenced by sampling variability, particularly when the number of above-threshold tornadoes is limited. Therefore, the cross-State performance comparisons should be interpreted with consideration of both the point estimates and their associated CIs, rather than relying solely on the observed performance values.

Figure 10 presents the permutation feature importance values obtained from the cross-State validation analysis. Consistent with the IA results, tornado path width remained the most influential predictor in nearly all States, followed by tornado path length as shown. The sine and cosine values of the bearing angle generally ranked third and fourth in importance, although their relative rankings vary among States.

One notable exception is Kansas, where tornado path length exhibits a slightly negative permutation feature importance value. A negative permutation importance indicates that randomly permuting the feature results in a small improvement in RF model performance, suggesting that the feature contributes little useful predictive information for this dataset. One possible explanation is that Kansas contained a number of EF3 tornadoes with relatively short path lengths, weakening the expected relationship between tornado path length and maximum DWS. Consequently, path length provided limited discriminatory power for distinguishing above-threshold and below-threshold tornadoes in Kansas.

Florida also exhibited distinctive behavior, with the sine and cosine values of the bearing angle having near-zero or slightly negative permutation importance values. These results suggest that tornado direction contributes little predictive information to the RF model in Florida. This may be related to the geographic characteristics of the Florida peninsula and the meteorological conditions that influence tornado development and movement within the State.

Overall, the feature importance results indicated that tornado path width and path length are the primary geometric characteristics to predict the maximum DWS among all examined geometric characteristics, while the contribution of tornado direction varies among States and is generally less influential. These findings further support the hypothesis that the relationship between tornado path geometry and maximum DWS exhibits State variability.

Different States 136 RF model results

Figure 11 summarizes the results of the State-specific 136 RF models and their independent validation datasets. Overall RF model performance varies among States, with accuracy and weighted F1 scores generally ranging from approximately 75%–95%. The corresponding validation results are similarly strong, with most validation datasets producing accuracy and weighted F1 scores near or above 85%. Recall scores for both classes generally range from 85% to 100%, with the exception of Texas, where recall for above-threshold tornadoes is comparatively lower.

The consistently strong performance observed across the State-specific RF models and their independent validation datasets (Figure 11) further demonstrates the robustness of the proposed RF modeling framework. The close agreement between testing and validation results suggests that the RF models are learning meaningful relationships between tornado path geometry and maximum DWS rather than memorizing the training data. The variation in RF model performance among States reflects differences in the geometric characteristics of below-threshold tornadoes relative to the population of above-threshold tornadoes used for RF model development. For example, IA below-threshold tornadoes appear to be more distinct from the national population of above-threshold tornadoes than those observed in Texas, Mississippi, or Tennessee. The validation results therefore indicate that meaningful geometric differences exist between above-threshold and below-threshold tornadoes within each State.

Figure 12 provides the feature importance values for the State-specific 136 RF models. Unlike the cross-State validation results, greater variability in feature importance is observed among the State-specific RF models. As shown in Figure 12, width was the most important predictor for Iowa, Missouri, and Tennessee, with path length ranking second. In contrast, path length was the most important predictor for Texas and Mississippi, while width ranks second. The sine and cosine values of the bearing angle generally remained the third and fourth most important features.

A particularly noteworthy result is observed for Tennessee, where the sine value of the bearing angle ranked as the third most important predictor and contributes at a level comparable to path length. This behavior is uncommon among the State-specific RF models and suggests that tornado direction may play a more significant role in distinguishing tornado intensity in Tennessee than in other States. The relatively high importance of the bearing-angle variable indicates that directional characteristics contain additional predictive information for Tennessee tornadoes that is not present to the same degree in most other States.

The variation in feature importance among the State-specific RF models further supports the need for multi-State regional analyses and State-specific RF model development. Although width and length consistently remained the dominant predictors, the relative contributions of the geometric variables varied from State to State, indicating that the relationship between tornado path geometry and maximum DWS is not uniform.

Relationship to structural and wind engineering

The structural engineering community uses fragility functions, or conditional probabilities, to predict the Degree of Damage to a given structural type for a given DWS. For example, tornado structural damage would involve the probability of Degree of Damage, for a building described by Damage Indicator 2, as a function of the estimated DWS. In this manner, estimates of maximum damage speed may be relevant even if the surface pressures would be better loading predictors. Using DWS exceedance predictions as shown here would be useful for calculating fragility values for various building types.

(Roueche et al., 2016) developed empirically based fragility functions for residential wind damage using logistic regression where the independent variable was estimated wind speed. (Rouech et al., 2018) described epistemic uncertainties derived from in situ damage estimates using fragility functions. (Roueche et al., 2020) examined the performance of a low-rise building to tornado-induced and straight-line wind loads using boundary layer wind tunnel data. The peak tornado-induced pressures were on average 13% higher than the equivalent straight line winds. Later on in 2020 (Jain et al., 2020) developed an engineering-based tornado damage assessment tool [ETDA] for assessing the tornado vulnerability of residential structures. The ETDA tool uses simplified generic low-rise building probabilistic models of tornado wind speed, wind pressure, atmospheric pressure drops, and models of the structural capabilities of the building components. It also includes a wind-borne debris impact, which has been shown to cause extensive damage. However, the simulations require extensive calculations for each location. Because of the difficulties in expanding logistic regression for characterizing tornado magnitude based on multiple variables such as length, width, geolocation, duration, month, etc., the present ML method for estimating the exceedance of DWS that embodies all of these variables would be an improvement.

Relationship to tornado event analysis

The ML models here have shown that all EF3 tornadoes are not identical; that is, they do not follow the same path, in accordance with statements by (Fujita, 1978). A tornado varies from location to location in ways that can be illustrated by the ML models. Including tornado event data from year to year is critically important for updating all prediction models. In addition, given that the DWS probabilities are geo-based, the risk to communities will be easier to assess versus using the tornado day process.

Advantages and disadvantages of data driven modeling for tornadoes

Advantages

The results of this study demonstrate the ability of data-driven models to identify relationships between tornado path characteristics and maximum DWS without requiring predefined assumptions regarding the functional relationship between variables. In particular, tornado path width consistently emerged as one of the most influential predictors across multiple State-specific models. This finding is consistent with previous observations that wider tornado paths are often associated with stronger tornadoes and greater damage footprints. However, the ML framework provides additional insight by quantifying the relative contribution of multiple geometric variables simultaneously rather than examining individual relationships independently.

The feature importance analysis also identified differences in the relationship between tornado geometry and damage intensity in different States. For example, path length generally showed a positive contribution to model performance in most States, consistent with previous observations that stronger tornadoes tend to produce longer damage paths. However, the Kansas model exhibited slightly negative permutation feature importance for path length. Examination of the Kansas validation data revealed several EF3 tornadoes with relatively short paths, suggesting that path length alone may not provide consistent information for distinguishing high-intensity tornadoes in this State. Rather than treating this result as a model limitation, the data-driven approach highlights an observation that may indicate State differences in tornado structure or limitations in current damage-based intensity characterization.

Disadvantages

The performance of data-driven tornado models remains strongly dependent on the availability and reliability of historical observations. This limitation is particularly important for severe tornado events. Although thousands of tornado records are available after the implementation of the EF scale, the number of EF3 and stronger tornadoes remains relatively small compared with weaker events. As a result, models trained on historical observations inherently contain fewer examples of the events that are most important for structural damage assessment. This challenge is especially significant for EF4 and EF5 tornadoes, where the limited number of observations restricts the ability of ML models to fully characterize the relationship between tornado geometry and extreme damage. Another important limitation is uncertainty associated with historical tornado damage records. Although extensive data cleaning procedures were performed in this study, tornado intensity estimates remain influenced by survey practices, available damage evidence, and changes in rating methodologies over time. This also influences the ML model performances.

Proposed future research directions

State-specific and multi-state RF models, and Cross-State generalizability

Future studies should extend the proposed framework to other tornado-prone States. The results presented in Section 3.3 demonstrate that the framework performs consistently across multiple States, suggesting that State-specific RF models can be developed for many States. The development of a library of State-specific RF models would allow for variations in different States.

The cross-State validation results demonstrate that tornado geometry–damage relationships vary among different States. The IA-based RF model performed well when applied to several neighboring States although performance decreased in other States where tornado characteristics and available observations differed. The grouped States of Illinois, Missouri, Nebraska, Wisconsin, and Minnesota exhibit performance metrics similar to those obtained for IA. This suggests a potential for developing multi-States RF models that combine tornado data from neighboring States with comparable climatological and geographic characteristics. In addition, States with fewer strong tornado events that are not enough to train the State-specific model can use this strategy to construct their RF models.

The uncertainty associated with the cross-validation results presented in Section 3.3 should also be considered when comparing model performance among States to demonstrate that tornado geometry–damage relationships among different States as well as developing multi-States RF models. Further research using additional State-specific models for cross-State validation and larger tornado datasets, including events from earlier periods is needed to further evaluate tornado geometry–damage relationships among different States and to assess the generalizability of the proposed modeling framework.

Integration of hazard model and fragility functions for tornado risk

A natural extension of the proposed ML framework is to integrate hazard predictions with State-specific empirically-based fragility functions developed from post-event damage data. (Roueche et al., 2016) developed empirical logistic (“S-curve”) models relating DWS to the probability of exceeding specific Degrees of Damage for individual Damage Indicators defined by the EF scale. These functions describe the conditional probability that a structure reaches or exceeds a specified damage state.

Using damage point data from DAT, similar functions could be developed for different building types and States using ML logistic regression. Although the underlying mathematical formulation remains logistic regression, a ML provides a data-driven approach in which model performance can be evaluated through independent testing, and established classification metrics. This enables both the predictive capability and uncertainty of the functions to be quantified. Integrating the proposed hazard model with these functions would provide a complete probabilistic framework for tornado risk assessment. For example, the RF model developed in this study can estimate the probability that an IA tornado exceeds a DWS of 150 mph (67.1 m/s). A corresponding function for one- and two-family residential buildings (Damage Indicator 2) can then estimate the probability that those buildings exceed Degree of Damage 4 when subjected to a 150-mph (67.1 m/s) DWS. Combining these conditional probabilities provides an estimate of the overall probability that a building experiences severe or life-threatening damage during a tornado event.

An additional advantage of this framework is that it enables State differences to be quantified directly from observed damage data. Buildings in different States are designed under different construction practices, building inventories, and code requirements. Consequently, these functions developed independently for different States are expected to exhibit different thresholds.

Such a framework would provide valuable information for engineers, emergency managers, and policymakers. As damaging tornadoes continue to occur outside the traditional Great Plains corridor, hazard and vulnerability assessments could support decisions regarding tornado-resistant design, emergency planning, and future building code development. By combining DWS exceedance probabilities with State-specific fragility functions, the proposed framework moves beyond tornado intensity prediction toward a comprehensive, performance-based design.

This study developed a ML framework to estimate tornado DWS using historical tornado damage records across the United States from 2007 to 2025, corresponding to the implementation of the EF scale. Tornado damage records were systematically collected and quality controlled to establish a reliable dataset for model development. ML approaches were investigated to identify effective predictors and modeling strategies for distinguishing tornadoes exceeding selected DWS thresholds.

The results demonstrate that data-driven models can identify complex relationships between tornado path characteristics and maximum DWS without requiring predefined assumptions regarding the functional relationships among variables. The developed State-specific RF models achieved strong predictive performance across the evaluated DWS thresholds, with high overall classification performance and recall for above-threshold tornadoes. Comparisons with alternative ML approaches further demonstrated the effectiveness of the RF framework for the imbalanced classification problem considered in this study. The cross-State validation results with 95% CIs also showed that models developed using tornadoes from one State can provide useful predictions for tornadoes in other States, although the level of performance varies among States and uncertainties remain. These results support the development of State-specific and multi-State models while emphasizing the importance of evaluating model transferability rather than assuming that a single model is universally applicable.

Feature importance analyses provided additional insights into the model results. Tornado path width was consistently identified as an important predictor across multiple State-specific models among all examined predictors, while path length also contributed to the predictions in many cases. The importance of directional features was generally smaller than that of the primary geometric characteristics.

The study demonstrated that RF outputs can be calibrated to provide State-level DWS exceedance probabilities. Independent temporal calibration validation using IA tornadoes from 2007 to 2014, which were not used for RF model development or calibration, provided an independent assessment of the complete RF prediction and calibration framework. The validation results demonstrated improved overall probabilistic performance after calibration, as reflected by reductions in Brier scores across the evaluated DWS thresholds, while also indicating threshold-specific variability in calibration performance. This variability may partly reflect the prediction performance of the underlying RF models when applied to the temporally independent dataset. These results provide evidence that the proposed framework can extend RF-based tornado classification to State-level exceedance probability estimation.

The study also highlights several challenges associated with ML-based tornado modeling. The limited number of high-intensity tornado observations, particularly EF4 and EF5 events, introduces uncertainty when modeling rare extreme events. In addition, historical tornado damage records contain uncertainties associated with survey methods, rating practices, and changes in data collection procedures over time. Cross-State validation further demonstrated that model performance varies among States, indicating that tornado geometry–damage relationships are not necessarily uniform across the country. Therefore, model predictions should be interpreted as data-driven estimates based on available observations rather than direct representations of tornado wind fields.

Overall, this study demonstrates the value of ML as a tool for analyzing complex tornado damage datasets and provides a framework for improving tornado intensity estimation. By combining data-driven modeling with physical analysis, future research may advance understanding of tornado damage behavior and support more reliable tornado risk assessment and resilient design practices.