Neural Networks Learn to Trust Their Data: Uncertainty-Weighted AI Predicts Asphalt Density

Every laboratory measurement in materials science carries a hidden burden: an uncertainty that tells engineers how much the result can be trusted. For decades, test standards have quantified this uncertainty through round-robin campaigns in which multiple laboratories measure identical samples and their spread is distilled into a reproducibility figure. Yet when machine learning models are trained to predict material properties, that hard-won knowledge is routinely discarded. A new study published in Case Studies in Construction Materials by Mohammad Abbas and Martins Zaumanis argues that this omission is a significant blind spot, and it offers an elegant fix: a neural network that weighs every training example according to how physically reliable its measurement actually is.

The researchers call their approach the Uncertainty-Weighted Neural Network, or UWNN. The idea is deceptively simple. In a conventional artificial neural network, the loss function that guides training is the mean squared error, which treats every data point as equally trustworthy. The UWNN replaces this with a weighted mean squared error in which each residual is divided by the reproducibility standard deviation defined in the relevant test standard. A result from a test known to be noisy therefore contributes less to the training signal than a result from a tightly controlled measurement. The network learns, in effect, to minimize its errors relative to what the measurement process itself can resolve, rather than chasing noise as if it were signal.

To demonstrate the concept, the team turned to asphalt pavement engineering, a field where bulk density is king. In the European Marshall framework, the bulk density of a compacted cylindrical specimen governs the air void content, the voids filled with bitumen, and the voids in the mineral aggregate — parameters that determine the optimum binder content and set the baseline for achieving adequate compaction in the field. Air void content in particular controls how a pavement resists rutting, moisture damage, and aging. As recycled materials such as Reclaimed Asphalt Pavement enter mixtures in growing proportions, precise density measurements become even more critical, because the inherent variability of reclaimed material must be tracked with confidence.

The evidence base for the study is unusually rich. The authors assembled 3,981 quality control test results for asphalt mixtures produced at plants in the Swiss cantons of Aargau and Zurich, with 720 records collected between 2017 and 2025 and 3,261 records between 2020 and 2024, all fully compliant with European standards. Specimens were Marshall-compacted with 50 blows per side, bulk density was measured by the saturated surface dry procedure, and maximum density followed the standard volumetric method. The dataset spans surface, binder, base, and foundation layers, and includes both unmodified and SBS polymer-modified binders. Bulk densities range from roughly 2.25 to 2.55 megagrams per cubic meter, a spread wide enough to let any model generalize beyond a narrow window of mixtures.

Before training anything, the researchers examined how the mixture composition features relate to bulk density. No single input dominated: reclaimed asphalt content correlated moderately at 0.59, binder content negatively at −0.55, and several mid-sized sieve fractions showed correlations around −0.5, while the finest fraction below 0.063 millimeters showed essentially none. This absence of a single strong predictor justified a multivariate machine learning approach. Missing sieve values were reconstructed using linear interpolation in log-log space, exploiting the fractal power law that particle size distributions in asphalt concrete follow, with a mean coefficient of determination of 0.991 for the reconstruction.

The architecture itself remains a standard feedforward network, tuned by Bayesian optimization over learning rate, weight decay, dropout, depth, width, and batch size. Strikingly, the optimal hyperparameters for the UWNN and the baseline ANN came out nearly identical — four hidden layers of sixteen neurons, a learning rate near 0.004, and a batch size of 32 — which allowed the authors to isolate the effect of the uncertainty weighting itself. On the full dataset, evaluated with ten-fold cross-validation and a held-out test set of 398 records, the UWNN delivered modest gains: a 4.7 percent reduction in the reproducibility-adjusted root mean squared error, a 2.4 percent improvement in the weighted RMSE, and slightly better coefficients of determination. Both models achieved prediction interval coverage close to the ideal 95 percent, meaning nearly all predictions fell within the reproducibility band that the test standard itself defines.

The more revealing finding was stability. Across cross-validation folds, the UWNN reduced the standard deviation of its performance metrics by roughly 20 percent compared with the ANN, indicating that its predictions were far less sensitive to which subset of data it happened to train on. The authors also extended the framework to air void content, propagating the bulk density uncertainty through the standard air voids equation using first-order uncertainty propagation. When judged against Swiss specification limits derived from EN 13108-1, the model classified compliance well for most mixture types, though performance varied: the AC T 22 H mixture showed the most balanced behavior, while AC T 22 S misclassified 12.4 percent of compliant samples, a caution the authors attach explicitly to predictions near specification boundaries.

The most dramatic results came from a stress test designed to mimic real-world data quality. The team trained both models on just 10 percent of the training pool — 358 records, closer to what a typical materials laboratory actually possesses — and deliberately corrupted a fraction of the training targets with Gaussian noise, targeting samples drawn from the quartile with the highest measurement uncertainty. Because the UWNN assigns these uncertain samples roughly half the weight of the others, the corrupted quartile contributed only about 15 percent of the total loss, while the ANN absorbed the noise at full strength. Across 25 independent training runs, the UWNN’s advantage grew with corruption: the improvement in the reproducibility-adjusted R-squared climbed from 5.6 percent with clean data to 30.5 percent at 20 percent corruption, a nearly six-fold gain, while error metrics more than doubled their advantage. At 10 and 20 percent corruption, paired t-tests and Wilcoxon signed-rank tests both confirmed the improvement was statistically significant.

The implications reach well beyond asphalt. Because the UWNN is a minimal modification of a classical neural network, it can be ported to virtually any materials testing domain where test standards define reproducibility — concrete strength, binder properties, soil mechanics, and beyond. The authors are careful about the limits: when uncertainty is constant across all observations, the weighted loss reduces to the ordinary loss, and when large, high-quality datasets exist, a conventional network is already near its ceiling. The framework pays off precisely where data are scarce, noisy, or partly unreliable — conditions that describe most real laboratories. The predictive accuracy also remains bounded by what the quality control records contain: aggregate shape, mineralogy, and binder properties are absent, which explains the small fraction of predictions falling outside the reproducibility band.

In an era when machine learning is being eagerly adopted across civil engineering, this study delivers a refreshingly sober message: the smartest model is not necessarily the one that fits the data hardest, but the one that knows how much each data point can be trusted. By writing the wisdom of the test standards directly into the training objective, the UWNN turns measurement uncertainty from an afterthought into an asset — and suggests that the future of AI in materials testing may depend less on bigger networks than on better-informed ones.

Subject of Research: Uncertainty-weighted neural networks for predicting asphalt mixture volumetric properties from quality control test data

Article Title: Uncertainty‑weighted neural networks in materials testing: Application to asphalt mixture volumetric properties

Article References: Abbas, M., & Zaumanis, M. (2026). Uncertainty‑weighted neural networks in materials testing: Application to asphalt mixture volumetric properties. Case Studies in Construction Materials, 25, Article e06613. https://doi.org/10.1016/j.cscm.2026.e06613

Image Credits: AI Generated

DOI: 10.1016/j.cscm.2026.e06613

Keywords: machine learning, neural networks, measurement uncertainty, asphalt mixture, bulk density, air voids, Marshall test, pavement engineering, quality control, reclaimed asphalt pavement, Bayesian optimization, Switzerland