According to WPB, researchers in China have developed a new artificial intelligence model that combines pavement damage mechanics with machine learning to predict asphalt deterioration under traffic loading, temperature variations, and freeze–thaw cycles. The study, published on November 7, 2026, evaluated 216 laboratory specimens alongside pavement condition, climate, and traffic data. The model achieved a coefficient of determination (R²) of 0.91 for rut depth prediction and 0.87 for cracking ratio, demonstrating strong predictive performance across the investigated conditions.
The research was conducted by scientists affiliated with Southeast University Chengxian College and China Design Group in Nanjing. It addresses a major challenge in pavement engineering: predicting how asphalt damage develops over time when mechanical loading and environmental conditions interact. Conventional prediction methods often rely on historical performance data or simplified engineering relationships, making it difficult to represent the combined effects of traffic, temperature, aging, and repeated environmental exposure.
To address these limitations, the researchers developed a multiscale framework connecting damage processes at microscopic, intermediate, and pavement structural levels. The approach examines how deterioration begins within asphalt materials, develops through interactions between aggregates and binder, and eventually appears as measurable pavement damage, including rutting and cracking.
The experimental program involved 216 independent specimens subjected to several types of laboratory testing. These included dynamic modulus measurements, four-point bending fatigue tests, wheel-tracking tests, and combined freeze–thaw and ultraviolet aging procedures. The researchers also used image-based reconstruction to examine the internal structure of asphalt mixtures and relate material-level deterioration to overall pavement behavior.
The prediction system integrates a gated recurrent neural network with an attention mechanism and equations describing physical damage evolution. This combination allows the model to learn complex patterns from experimental and pavement data while following established mechanical principles. Physical constraints were incorporated into the training process to prevent predictions from developing unrealistic damage trends.
The results showed strong performance in predicting two important indicators of pavement deterioration. For rut depth, the model achieved an average R² of 0.91, with a variation of ±0.01 across five training runs. For cracking ratio, the corresponding result was 0.87 ±0.02. These values measure the statistical agreement between predictions and observed outcomes; they do not represent a guaranteed 91% or 87% accuracy rate for every road.
One of the most significant improvements was the model's ability to maintain physically consistent predictions. Introducing damage-mechanics constraints increased the consistency of extrapolated deterioration trends from 0.79 to 0.93. At the same time, violations of the expected direction of damage progression fell from 11.6% to just 0.3%. This improvement is particularly important when predicting pavement behavior beyond the conditions directly represented in the training data.
The researchers also evaluated the reliability of the model's uncertainty estimates. The actual coverage of its nominal 90% prediction interval reached 89.4%, indicating that the calculated uncertainty range closely matched the intended statistical coverage during testing. This capability is relevant to infrastructure management, where maintenance decisions depend not only on a predicted damage level but also on the reliability of that prediction.
Temperature emerged as another important factor influencing asphalt deterioration. The study identified nonlinear damage behavior across three aggregate gradations, with stronger cross-scale effects occurring below freezing and at temperatures above approximately 45°C. The magnitude of these effects varied between mixture gradations, illustrating the importance of material composition when evaluating pavement performance under different climatic conditions.
To assess how well the model performed beyond its training data, the researchers separated the testing information by road section, observation period, and geographical region. This approach was designed to provide a more demanding evaluation than randomly dividing closely related observations between training and testing datasets.
The research also connects damage prediction with infrastructure resilience. The proposed framework distinguishes between pavement resistance to deterioration, recovery following maintenance, and adaptation to changing operating conditions. These dimensions could support more detailed comparisons of maintenance strategies and help road authorities identify where preventive intervention may provide greater long-term benefits.
For asphalt producers, pavement engineers, and road construction authorities, the technology offers a potential improvement in evaluating pavement performance throughout its service life. More reliable predictions of rutting and cracking could support material selection, pavement design, inspection planning, and the scheduling of maintenance activities.
The development is also relevant to the bitumen industry because binder properties influence asphalt mixtures' resistance to deformation, fatigue, and environmental aging. Models that integrate mechanical behavior with real operating conditions could help researchers assess how different asphalt formulations respond to traffic and temperature. The technology is a prediction and engineering assessment tool, however, rather than a newly developed bitumen additive or binder modification.
The researchers have made the study's data, code, and model configuration files available to support reproducibility. The reported performance reflects the materials, datasets, and validation procedures investigated in the research. Wider application will require additional testing across different pavement structures, climatic regions, traffic conditions, and long-term operating environments.
The study represents an advance in combining artificial intelligence with established principles of pavement mechanics. By improving both predictive performance and physical consistency, the new framework offers a foundation for more reliable asphalt condition assessments and better-informed maintenance planning, with potential applications in extending pavement service life and improving infrastructure management.
By WPB
Artificial Intelligence in Asphalt, Asphalt Pavement Prediction, AI Pavement Monitoring, Machine Learning Asphalt, Asphalt Rutting Prediction, Asphalt Cracking Prediction, Pavement Damage Mechanics, Physics-Informed AI, Pavement Life-Cycle Performance, Asphalt Durability, Freeze-Thaw Damage, Road Maintenance Technology, Predictive Pavement Modeling, Smart Road Infrastructure, Asphalt Engineering, Bitumen Technology, Pavement Resilience
