The stakes are enormous. Overloaded and improperly distributed vehicle loads accelerate pavement deterioration, deepen rutting and cracking, and impose punishing fatigue cycles on bridge components. When more than twelve thousand vehicle load records were analyzed in earlier foundational work, researchers realized that a truck’s structural configuration carried much of the information needed to infer its weight—a premise this new study pushes to its practical limit. Modern pavement and bridge design codes rest on strict federal axle-load and gross-weight limits precisely because overweight traffic erodes the safety margins engineers assume when they design these assets. Studies using Weigh-In-Motion data have shown that even modest increases in the share of overloaded vehicles can sharply shorten the fatigue life of flexible pavements, and mechanistic-empirical analyses confirm that dynamic axle-load spectra arising from road unevenness intensify both rutting and fatigue cracking. The economic toll lands squarely on transportation agencies and road users alike.

To understand the solution, it helps to understand the technology gap. Weigh-In-Motion systems—embedded sensors such as piezoelectric cables, bending plates, and load cells—record axle loads, vehicle classes, and gross weights without interrupting traffic flow, and they have proven invaluable for freight analysis. But their installation and maintenance costs are substantial, limiting deployment to a handful of carefully chosen highway segments. By contrast, Automated Traffic Recorders are far more widely distributed and routinely collect variables such as vehicle class, number of axles, and speed. The insight behind the new research is deceptively simple: if a machine-learning model can learn the relationship between those commonly available attributes and gross vehicle weight from data-rich WIM stations, that same model can be exported to ATR-only corridors, effectively weighing trucks on roads where no scales exist.

The research team, led by Eric Osei with colleagues including Henry Sawaki, Arthur Mukwaya, Judith Mwakalonge, Gurcan Comert, Saidi Siuhi, Awadh Simba, and Japhet Felician, built their framework on 2021 WIM datasets from three geographically distinct sites: the Brooklyn–Queens Expressway at Pearl Street in New York, Van Nuys along Route 405 in Los Angeles County, and Wise County along East State Highway 114 in Texas. Together the three datasets comprised roughly ten million vehicle records. The team restricted their analysis to heavy-duty freight vehicles—FHWA Classes 4 through 13—to ensure consistent definitions across datasets, converted all weights from kips to kilonewtons using the standard factor of 1 kip equaling 4.44822 kN, and rigorously cleaned the data by discarding physically impossible records such as non-positive gross weights, invalid axle readings, and vehicles with fewer than two axles. Critically, they deliberately excluded the axle-level measurements that only WIM systems provide, restricting their candidate inputs to variables any Automated Traffic Recorder could supply: vehicle class, number of axles, and temporal attributes such as hour, day, and month.

Feature selection proved to be one of the study’s most consequential steps. The researchers screened candidate variables with correlation analysis and then validated their choices using SHAP—SHapley Additive exPlanations—a game-theoretic attribution technique that decomposes a prediction into contributions from each input feature. Across all three sites, the results were unambiguous: vehicle class and number of axles dominated gross vehicle weight prediction, while temporal variables such as hour, day, and month contributed negligibly. This makes physical sense, since vehicle class encodes the regulatory and structural classification that inherently defines a truck’s permissible weight range, and axle count serves as a direct mechanical indicator of load distribution and carrying capacity. The verdict allowed the team to collapse the feature set to just two variables, a deliberately compact design that maximizes real-world deployability while keeping the models interpretable and minimizing the risk of data leakage.

Six modeling approaches were then pitted against one another: a Lookup Table baseline that estimates weight as the conditional mean for each class-and-axle combination, Multiple Linear Regression, Class-Specific Regression that fits separate relationships within each FHWA class, a Generalized Additive Model capturing smooth nonlinear effects, Random Forest, and Extreme Gradient Boosting. Random Forest aggregates predictions across many decorrelated decision trees, each grown on a bootstrap sample with random feature subsets, reducing variance and improving robustness. XGBoost, by contrast, builds an additive ensemble stage-wise, fitting each new tree to the residual errors of the current ensemble while controlling complexity through built-in regularization. The evaluation protocol was unusually rigorous for this field: repeated random-split validation with five seeds, five chronological train-test configurations simulating real deployment on future traffic, and performance summarized as means with standard deviations and 95 percent confidence intervals.

The results delivered a striking trade-off between accuracy and portability. Under random-split validation, XGBoost achieved the highest in-site accuracy, reaching a coefficient of determination of 0.84 on the New York data, 0.83 in California, and 0.78 in Texas, with Random Forest close behind at 0.81, 0.80, and 0.75 respectively—dramatic improvements of roughly 40 to 50 percent over the Lookup Table baseline. Under the harsher time-based validation, where models trained on earlier months had to predict later ones, XGBoost still led with R² values of 0.80, 0.79, and 0.75, and paired t-tests confirmed its edge over Random Forest was statistically significant at every site, with p-values between 0.031 and 0.042. But when the researchers transferred models across sites without retraining—six directional source-target combinations in total—the hierarchy flipped. Random Forest, with its bootstrap-aggregation architecture that reduces variance and tolerates distributional differences, emerged as the most transferable model, retaining approximately 92 percent of its in-site predictive capability. XGBoost retained about 83 percent, an outcome the authors attribute to boosting’s tendency to fit residual errors that encode site-specific relationships, making it more sensitive to local data characteristics.

This cross-site analysis gave rise to the study’s signature contribution: the Transferability Index, defined as the ratio of cross-site R² to in-site R², converting what had long been a qualitative observation—that models do not travel well—into a measurable, comparable quantity. Random Forest posted Transferability Index values between 0.89 and 0.95 across all six transfer scenarios, with out-of-site R² values from 0.65 to 0.71, while XGBoost ranged from 0.80 to 0.85. Paired t-tests on the transfer results confirmed that Random Forest’s generalization advantage was statistically significant in every scenario, with p-values from 0.026 to 0.042. The baselines told a cautionary tale: the Lookup Table collapsed to out-of-site R² values as low as 0.16 with errors approaching 99 kilonewtons, while Multiple Linear Regression managed Transferability Index values near 0.67. The practical implication is clear—XGBoost is the instrument of choice when maximizing accuracy at a known location, but Random Forest is the better vehicle when a model must be deployed to unseen corridors.

What makes the work resonate beyond pavement engineering is the way it stitches together three previously separate strands of research. Attribute-based weight prediction had been demonstrated before, low-cost inference on unmonitored road segments was well established for traffic volumes—neural networks, kriging, convolutional networks, and quantile random forests have all been used to estimate annual average daily traffic where no permanent counters exist—and transfer learning frameworks had been applied to cross-city flow prediction. But no prior study had combined all three: gross vehicle weight estimation from a minimal, universally available feature set, evaluated across multiple geographically distinct stations, with transferability quantified through a statistically tested protocol. The framework also formalizes validation practices that transportation machine learning has often neglected, including leakage-free feature handling and blocked or grouped validation schemes appropriate for data with temporal and spatial structure.

The researchers are candid that all models lost some performance when transferred, reflecting genuine differences in traffic composition, freight patterns, fleet characteristics, and operational conditions between sites—TI values below unity make that loss explicit rather than hiding it. Still, the demonstration that two simple attributes can carry a Random Forest model across state lines with most of its predictive power intact represents a meaningful step toward affordable, network-wide weight monitoring. For highway agencies wrestling with premature pavement failure and accelerated bridge deterioration, the prospect of extending weight estimation to non-coverage roads using infrastructure they already own is the kind of quiet revolution that rarely makes headlines but reshapes how roads get built, maintained, and funded. As freight demand continues its relentless growth, the trucks rumbling down America’s unscaled highways may soon be weighed by nothing more than algorithms—and the roads will be stronger for it.