The stakes are enormous. In 2024, the United States experienced 27 separate billion-dollar weather and climate disasters, and tornadoes and thunderstorms accounted for more than half of them. Although violent tornadoes rated EF3 and above represent a small fraction of all occurrences, they dominate economic losses, infrastructure disruption and threats to life. Wind engineers, unlike meteorologists issuing real-time warnings, need long-term probabilistic estimates of peak wind speeds to inform design standards for buildings and infrastructure, and the broad EF categories obscure exactly the information they need most.

The core innovation of the study lies in what it predicts. Rather than assigning tornadoes to EF categories, the researchers model the probability that a tornado’s maximum damage wind speed (DWS) exceeds specific thresholds: 136, 140, 145, 150, 155, 160 and 165 miles per hour, all within the EF3 range of 136 to 165 mph. This range matters because it is a transition zone. Most structures built to current codes are expected to withstand EF2-and-below winds of 135 mph or lower, while EF4 and EF5 events above 166 mph often produce near-total destruction. The EF3 band, in which tornadoes cause a broad spectrum of severe but not necessarily complete structural damage, is where prediction offers the greatest engineering value, and EF3 tornadoes occur far more frequently than their stronger counterparts, providing more data for model development.

Building a trustworthy dataset was the study’s crucial first step. The team quality-controlled and cross-validated records from three federal sources: the Storm Events Database maintained by the National Centers for Environmental Information, the Damage Assessment Toolkit (DAT), and the Storm Prediction Center (SPC) tornado database. The DAT, which provides the most comprehensive post-event damage information, served as the primary source. Records with missing, inconsistent or duplicated values were flagged and investigated rather than simply deleted, and corrections were made only when supported by at least two independent sources, including National Weather Service and NOAA field survey reports. The result is a curated historical dataset of more than 630 authenticated EF3 through EF5 tornadoes in the United States from 2007, when the EF scale was introduced, through 2025.

With the data curated, the researchers turned to a Random Forest (RF) classifier, an ensemble method that constructs many decision trees and aggregates their predictions to improve accuracy and reduce overfitting. The model uses just four geometric features of the tornado path: track length, maximum width, and the sine and cosine components of the path bearing angle. The authors deliberately minimized the feature set because earlier studies showed that most variables in tornado databases contribute little predictive power, and because correlated variables such as duration, which tracks closely with path length, introduce redundancy. The bearing angle, a circular variable where 0 and 360 degrees mean the same direction, was decomposed into sine and cosine components to avoid discontinuities that would confuse the algorithm.

Model development centered on Iowa, home to the National Testing Facility for Enhancing Wind Resiliency of Infrastructure in Tornado-Downburst-Gust Front Events (NEWRITE) project and part of the midwestern tornado alley. Because violent tornadoes are rare, only 10 EF5 tornadoes occurred nationwide between 2007 and 2025, the team had to confront severe class imbalance. They enriched the training data with above-threshold tornadoes from across the country and applied the Synthetic Minority Over-sampling Technique (SMOTE) where necessary. Because this rebalancing inflates the apparent probability of extreme events, the researchers calibrated the raw RF outputs using logistic regression, following a method proposed by Dankowski and Ziegler, which produced closer agreement with observed Iowa exceedance rates than traditional Bayes prior correction.

The results are striking. Across all seven DWS thresholds, the Iowa RF models and their independent validation datasets consistently achieved accuracy and weighted F1 scores above 85 percent, with recall scores for both above-threshold and below-threshold tornadoes generally exceeding 85 percent. Permutation feature importance analysis revealed that tornado path width was by far the most influential predictor, followed by path length, with the directional variables ranking third and fourth. This makes physical sense: stronger tornadoes carve wider damage paths, and the relationship proved more direct than the link between intensity and path length, since some EF3 tornadoes leave shorter tracks than weaker storms. Comparisons with logistic regression and XGBoost baseline models showed broadly comparable overall performance, but the RF models maintained consistently higher recall for the above-threshold class, the metric that matters most for hazard assessment, where missing a dangerous tornado is far worse than a false alarm.

Perhaps the most consequential finding concerns geography. When the Iowa 136 mph model was applied to other states, performance varied substantially. Neighboring states such as Illinois, Missouri, Nebraska, Wisconsin and Minnesota performed comparably to Iowa, with Illinois reaching 89 percent accuracy, but Kansas, the Dakotas and Florida showed lower recall for above-threshold tornadoes, while Mississippi and Tennessee caught dangerous tornadoes well but misclassified weaker ones. In Kansas, path length even showed slightly negative feature importance, apparently because several EF3 tornadoes there had unusually short tracks. State-specific models developed for Missouri, Mississippi, Tennessee and Texas generally achieved accuracy between 75 and 95 percent, suggesting that tornado geometry-damage relationships are not uniform across the country and that state-specific or regional models may capture local variability better than any universal model.

The framework’s engineering implications extend well beyond classification. Structural engineers use fragility functions, conditional probabilities that a given building type exceeds a specific degree of damage at a given wind speed, to assess tornado risk. By combining the RF model’s DWS exceedance probabilities with state-specific fragility functions, the approach could estimate the overall probability that a residential building suffers severe or life-threatening damage in a tornado, moving the field toward performance-based design. The authors note that previous engineering-based damage assessment tools require extensive calculations for each location, whereas the machine learning framework embodies multiple variables at once and can be updated as new event data arrive each year.

The study also candidly acknowledges its limits. Only 10 EF5 tornadoes occurred in the study period, so the rarest and most destructive events remain poorly represented, and historical damage records carry uncertainties from survey practices, varying state building codes and evolving rating methodologies. The authors caution that model predictions should be interpreted as data-driven estimates rather than direct representations of tornado wind fields, especially since surface wind speeds during tornadoes remain nearly impossible to measure directly. Still, by demonstrating that a tornado’s path geometry, above all its width, carries quantifiable information about its destructive potential, the research opens a practical path toward probabilistic tornado hazard maps, fragility analysis and, ultimately, more resilient building design in the communities that need it most.

Subject of Research: Machine learning prediction of tornado damage wind speed exceedance from tornado path geometry

Article Title: Tornado damage prediction using machine learning

Article References: Tornado damage prediction using machine learning. (n.d.). https://doi.org/10.3389/fbuil.2026.1971791

Image Credits: AI Generated

DOI: 10.3389/fbuil.2026.1971791

Keywords: tornado, machine learning, random forest, Enhanced Fujita scale, damage wind speed, wind engineering, path width, Iowa, hazard assessment, fragility functions, Storm Prediction Center, structural design