According to WPB, artificial intelligence is moving one step closer to predicting how long an asphalt mixture can resist repeated traffic loading before fatigue cracking becomes a serious problem. A new study published online on August 5, 2026, introduces a model called DM-PINN-FP — Dual-Mode Physics-Informed Neural Network Fatigue Prediction — designed specifically to estimate the fatigue performance of asphalt mixtures. Unlike a conventional AI model that mainly searches for patterns in past data, this system is also trained to respect what engineers already know about how asphalt actually cracks.
That difference may sound technical, but the basic idea is straightforward. A conventional neural network can be given laboratory results from hundreds or thousands of asphalt specimens and asked to learn which combinations of temperature, loading and material properties are associated with longer or shorter fatigue life. If enough good data are available, that approach can work well. The problem is that asphalt fatigue testing takes time and money, while the performance of a mixture can change with temperature, loading conditions, binder characteristics, aggregate interaction and other variables.
The new approach attempts to give AI something more than a spreadsheet of previous results. It builds engineering knowledge about damage behavior into the learning process itself.
The researchers developed the DM-PINN-FP model around what they describe as the dual nature of fatigue damage in asphalt mixtures. Instead of treating cracking as one single type of failure, the model separates total fatigue damage into two components: cohesive damage inside the asphalt mastic and adhesive damage at the interface between the mastic and aggregate.
For people working with asphalt rather than AI, this distinction is important.
A road can crack because the binder-rich material within the mixture gradually loses its ability to withstand repeated loading. But cracks can also develop where the asphalt material bonds to aggregate particles. Those are different failure mechanisms, and changes in binder, aggregate, aging, temperature or mixture design may change which mechanism becomes more important.
The new model introduces a Cracking Preference Index, or CPI, into the fatigue-prediction process. The study links this index with material strength and incorporates it into the damage model used by the neural network. In practical terms, the CPI gives the system information about the mixture’s tendency toward different cracking mechanisms instead of asking the AI to discover everything from statistical relationships alone.
That is why the term “physics-informed” matters. The model is still learning from experimental data. It is not replacing pavement mechanics with a computer algorithm. But during training, its predictions are constrained by physical relationships that describe how damage should develop. If the network begins producing an answer that fits the training data statistically but behaves unrealistically from a materials perspective, the physics component is intended to push the solution toward a more reasonable result.
This addresses one of the biggest concerns surrounding purely data-driven AI in engineering. A neural network can sometimes perform extremely well on conditions similar to the data it has already seen but become unreliable when asked to predict something outside that range. For a music recommendation system, a poor prediction is inconvenient. For a pavement mixture expected to survive years of truck traffic, the consequences are considerably more expensive.
The researchers tested this issue by withholding one combined condition — 25°C and 400 microstrain — from the training data and then examining how the model behaved when predicting that unseen condition. The purpose was to evaluate whether the model could do more than memorize relationships already contained in its training set.
The study reports that DM-PINN-FP reduced the cumulative-error problem associated with standard neural-network prediction and performed better than a purely data-driven artificial neural network. In the model comparison presented by the researchers, DM-PINN-FP produced the lowest root mean square error and mean absolute error among the evaluated approaches.
For the asphalt industry, the important point is not the names of those statistical measures. It is that adding physical knowledge improved the model’s ability to keep its predictions closer to the measured fatigue behavior.
This research is also part of a larger change already underway in pavement engineering. A 2024 study introduced another physics-informed neural-network approach, known as PINN-AFP, to predict asphalt-mixture fatigue behavior using the Viscoelastic Continuum Damage approach. That system was designed to reconstruct a complete damage characteristic curve from a relatively small amount of early fatigue-test information. In its AC-25 case study, the model reported an average fatigue-life prediction error of about 5.2%, outperforming the machine-learning and deep-learning methods used for comparison.
Physics-guided AI has also been applied to asphalt rutting. Another published model combined conventional neural-network learning with physical information specifically to improve not only accuracy but also stability and engineering rationality. The researchers found that adding physics significantly improved the model’s behavior while maintaining high prediction accuracy.
DM-PINN-FP pushes this direction further by focusing on the competing cracking mechanisms inside the asphalt mixture.
That could eventually matter to mixture designers. Today, developing or approving a mixture may require several laboratory tests covering cracking, rutting, moisture sensitivity and other performance risks. Fatigue testing can be especially demanding because specimens must undergo repeated loading until enough damage develops to characterize their behavior.
A reliable predictive system would not necessarily eliminate those tests. More realistically, it could help laboratories decide which mixtures need extensive testing and which combinations can be screened more quickly.
Imagine a producer evaluating several binder contents, two aggregate sources, different binder grades and a recycling additive. Testing every possible combination under every relevant loading and temperature condition quickly becomes expensive.
A physics-informed model could eventually act as a screening layer. Laboratory results from a smaller selection of mixtures could feed the model. The model could then estimate the fatigue behavior of other combinations and identify which ones appear promising enough to justify full performance testing.
The laboratory would still make the final verification, but the amount of trial-and-error work could potentially decrease. This is where the technology becomes particularly relevant to Balanced Mix Design.
Balanced Mix Design is increasingly being used to move asphalt specifications beyond simply meeting volumetric requirements. The approach uses performance tests to consider several possible forms of pavement distress and aims to develop mixtures suitable for their specific traffic, climate, aging and pavement conditions. The U.S. Federal Highway Administration continues to support state implementation of BMD and performance-based asphalt testing.
Fatigue cracking is already part of that transition. FHWA tools use cyclic fatigue testing to characterize cracking behavior and provide inputs for pavement-performance analysis. The agency has also developed fatigue-related parameters that can be used in performance-engineered mixture design.
AI models such as DM-PINN-FP therefore do not have to create an entirely new industry process. In the longer term, they could potentially fit inside a process that is already becoming more performance-oriented.
A contractor or agency could start with candidate aggregate, binder and recycled-material combinations. Laboratory performance data could establish the physical behavior of several reference mixtures. A physics-informed model could then help estimate fatigue life under additional conditions before the most promising designs move into full testing and field validation.
The same concept could eventually influence binder selection. Binder is only one part of fatigue performance, and the new study predicts mixture behavior rather than offering a simple binder-ranking tool. Aggregate structure, adhesion and mixture characteristics remain important. But because the model explicitly distinguishes cohesive and adhesive damage, future versions could help engineers understand whether changing a binder is likely to improve the part of the system that is actually controlling failure.
That could be more useful than simply asking whether one binder is “better” than another.
A polymer-modified binder, for example, may improve one damage mechanism while a weak binder–aggregate interface continues to limit the mixture. If a digital model can identify which failure mode is dominating, material selection becomes a more targeted decision.
The technology may also have value as recycled asphalt contents increase.
Reclaimed asphalt pavement, recycling agents, polymer modifiers and other additives create more possible mixture combinations. That flexibility is valuable, but it also increases the number of variables laboratories must evaluate. Performance-based design becomes difficult to scale if every change requires a completely new series of long-duration tests.
Physics-informed AI could eventually make that process more manageable by using limited test results together with established mechanics to explore a larger design space. There is an important limitation, however: DM-PINN-FP is research, not a ready-to-use commercial pavement specification.
The publication demonstrates a modeling approach under the conditions and datasets used by the researchers. Before a road agency or producer could rely on such a model for routine mixture approval, it would need much broader validation across different binders, aggregate sources, mixture types, aging conditions, temperatures, loading frequencies, recycled-material contents and laboratory systems.
Field validation would be even more important. Predicting laboratory fatigue life and predicting when cracking appears on a highway are related problems, but they are not identical. Real pavements experience changing temperatures, moisture, aging, different axle loads, construction variability and structural conditions that a laboratory specimen cannot completely reproduce.
AI should therefore be viewed as a tool for improving engineering decisions, not as a replacement for testing or engineering judgment. The real opportunity may come when models like this are connected with digital pavement design.
A future system could combine mixture-test data, binder characteristics, traffic loading, climate information and pavement structure in one digital workflow. Instead of asking only, “Did this mix pass the fatigue test?”, engineers could ask how that mixture is expected to behave under a specific traffic level, temperature range and pavement structure — and compare alternatives before construction begins. That would move AI from simply recognizing pavement distress after it appears toward helping design mixtures before the distress occurs.
For bitumen producers, the development is worth following for the same reason. If asphalt specifications become increasingly performance-based, binder sales may gradually become more connected to predicted pavement behavior rather than traditional grade alone. Suppliers could face greater demand for reliable rheological, aging, adhesion and performance data that can feed digital mixture-design systems.
In other words, AI would not make binder quality less important. It could make the relationship between binder properties and actual pavement performance more visible. The immediate impact on the commercial bitumen market will be limited. DM-PINN-FP is not going to change global bitumen demand or specifications overnight. But technology shifts in pavement engineering often begin years before they become procurement requirements.
Balanced Mix Design is already expanding the use of performance testing. Digital pavement tools are becoming more capable. Physics-informed machine learning is now being applied to fatigue, rutting and other pavement-performance problems. The pieces are beginning to connect.
The significance of the August 5 study is therefore not simply that another AI model achieved lower prediction error. It shows where asphalt engineering may be heading: AI systems that do not just learn what happened in previous tests, but also understand some of the physical reasons why the material fails. If that approach survives wider laboratory and field validation, fatigue prediction could eventually become faster, more targeted and less dependent on testing every possible mixture under every possible condition.
The road industry is still a long way from allowing an algorithm to approve an asphalt mixture on its own. But a model that can combine test data with the physics of cracking could become a valuable second set of eyes for laboratories, producers and pavement designers. For an industry trying to build longer-lasting roads while controlling testing costs, that is the part of this technology worth watching.
By WPB
News, Bitumen, Artificial Intelligence, Asphalt Technology, Fatigue Cracking, Physics-Informed AI, Balanced Mix Design, Pavement Design, Asphalt Binder, Road Technology
