🤖 machine learning

Predicting the fatigue life of asphalt concrete using neural networks

This study utilizes artificial neural networks with a mean square logarithmic error loss function to accurately predict the fatigue life of asphalt concrete based on strain, binder content, and air-void content, demonstrating that higher binder content significantly improves durability while providing open-source tools for future research.

Original authors: Jakub Houlík, Jan Valentin, Václav Nežerka

Published 2026-02-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jakub Houlík, Jan Valentin, Václav Nežerka

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine asphalt concrete (the stuff roads are made of) as a giant, complex cake. The "ingredients" are crushed stones and a sticky, tar-like glue called bitumen. Just like a cake can crack if you keep bending it back and forth, roads crack under the weight of millions of cars. This process is called "fatigue," and figuring out how long the road will last before it cracks is a huge headache for engineers.

Traditionally, to find out how long a road will last, engineers have to bake little test cakes in a lab and bend them over and over again. This is slow, expensive, and takes weeks.

This paper is about teaching a computer (specifically a "Neural Network," which is a type of AI that learns like a human brain) to predict how long these road "cakes" will last without needing to wait for the slow lab tests.

Here is the breakdown of what they did, using simple analogies:

1. The Ingredients Matter Most

The researchers focused on three main "ingredients" that affect the road's life:

Strain Level: How hard the road is being bent (like how much you flex a paperclip before it snaps).
Binder Content: How much of the sticky glue (bitumen) is in the mix.
Air Voids: How many tiny air bubbles are trapped inside the cake.

They gathered data from four different scientific studies, creating a "recipe book" of 245 different road samples.

2. Cleaning the Data (The "Box-Cox" Magic)

The data they collected was messy. Most roads lasted a "normal" amount of time, but a few lasted way longer, creating a long, lopsided tail in the data (like a few people in a room being giants while everyone else is average).

To fix this, they used a mathematical trick called a Box-Cox transformation. Think of this as a blender that smooths out the lumpy data so the computer can digest it better. They also checked for "outliers" (weird, broken data points) using a "Z-score," which is like a metal detector for bad data. Fortunately, their data was clean, so they didn't have to throw anything away.

3. Training the AI Brain

They built a digital brain (a Neural Network) to learn the relationship between the ingredients and the road's life.

The Architecture: They tried different shapes for this brain. They found that a brain with two hidden layers (two levels of thinking) and 200 neurons (200 little thinkers) in each layer worked best.
The Activation Function (ReLU): This is like the brain's "on/off" switch. They found that the "ReLU" switch was the best at keeping the brain active and learning, whereas other switches made the brain go to sleep or act too simply.
The Optimizer (RMSprop): This is the teacher guiding the brain. They found that the "RMSprop" teacher was the most stable and effective at correcting mistakes.

4. Two Different Ways to Measure Success

Because road life varies so wildly (some last a few thousand cycles, others millions), they tried two different ways to grade the AI's homework:

LMSE (Linear): This method cares a lot about getting the big numbers right. It's like grading a test where getting the hardest questions right gives you the most points. This method was very accurate overall.
LMSLE (Logarithmic): This method cares more about getting the small numbers right. It's like grading a test where you get extra points for not missing the easy questions. This was better at predicting the life of roads that don't last very long.

The Result: The AI got incredibly good at predicting the future. When tested on new data, it achieved a score (called R²) of 0.967 using the linear method. That's like getting an A+ on a very difficult exam.

5. What the AI Learned (The "Secret Sauce")

By looking at how the AI made its predictions, the researchers discovered some clear rules:

More Glue is Good: Roads with more bitumen (binder) tend to last longer. The glue holds the stones together and keeps them flexible.
Fewer Bubbles is Good: Roads with fewer air bubbles (air voids) last longer. Too many bubbles make the road weak and prone to cracking.
The Sweet Spot: The AI suggested that for a road to be truly durable, it should have at least 5.5% binder. However, they noted that this might be too much for the bottom layers of a road, where less glue is usually used to prevent the road from getting too soft in the heat.

6. The Catch (Limitations)

The authors are honest about the limits of their work:

The Recipe Book is Small: They only had 245 samples. While that's a lot for this specific type of test, it's small for a computer to learn everything.
Bias: Most of their data came from just a few studies. The AI might be "biased" toward those specific types of roads and might not be perfect for totally different road types it hasn't seen before.
No Crystal Ball: The AI is great at predicting roads that look like the ones in the training data. If you ask it about a road made of completely different materials, it might guess wrong.

The Bottom Line

This paper shows that we can use a smart computer program to guess how long a road will last based on its recipe (how much glue and how many air bubbles it has). This could save engineers a massive amount of time and money by reducing the need for slow, repetitive lab tests. The researchers even put their code and data online so others can check their work or try to improve the "recipe."

Technical Summary: Predicting the Fatigue Life of Asphalt Concrete Using Neural Networks

Problem Statement
The durability of asphalt concrete (AC) is fundamentally linked to its fatigue life, a parameter that dictates maintenance costs and structural longevity. Traditional assessment methods, such as the four-point bending (4PB) test, are resource-intensive and time-consuming, often requiring weeks for specimen preparation and curing, followed by hours or days of testing. While machine learning (ML) offers a pathway to predict fatigue life using existing data, previous applications of artificial neural networks (ANNs) in this domain have often lacked detailed architectural descriptions or failed to utilize extensive, heterogeneous datasets. Furthermore, the inherent variability of fatigue life data, which often follows a logarithmic distribution with significant skewness, poses challenges for standard regression models.

Methodology
The study employed ANNs to predict the number of cycles to failure ( $N_f$ ) based on three primary input variables: strain level ( $\varepsilon$ ), binder content, and air-void content.

Data Collection and Preprocessing: The dataset comprised 245 samples extracted from four peer-reviewed scientific publications, all utilizing 4PB tests at a constant strain rate, 10 Hz frequency, and temperatures of 20°C (or 70°F). To address the significant skewness and long-tail distribution of the original $N_f$ data, a Box–Cox transformation was applied with an optimal parameter $\lambda = -0.09$ . This normalization was validated using Shapiro–Wilk tests and Q-Q plots. Outlier detection was performed using both standard Z-scores and modified Z-scores (based on median absolute deviation); no outliers were identified, allowing the retention of the full dataset.
Model Architecture and Training: The ANN was developed using the Keras library (sequential model). Input variables were normalized using MinMaxScaler. The model architecture was optimized through a systematic hyperparameter search, evaluating the number of hidden layers ( $n_h$ ), neurons per layer ( $n_{neur,h}$ ), activation functions, and optimizers. The final selected architecture featured two hidden layers with 200 neurons each, utilizing the ReLU activation function and the RMSprop optimizer.
Loss Functions and Validation: To accommodate the wide range of fatigue life data, two distinct loss functions were employed: Mean Square Error (LMSE) and Mean Square Logarithmic Error (LMSLE). The dataset was split into 75% for training (further divided into 80% training and 20% validation) and 25% for testing. A four-fold cross-validation approach was utilized to ensure robustness and minimize overfitting. Model performance was evaluated using the coefficient of determination ( $R^2$ ), adjusted $R^2$ , and the Scatter Index (SI).

Key Results

Hyperparameter Optimization: The configuration of $n_h = 2$ and $n_{neur,h} = 200$ yielded the highest performance. ReLU consistently outperformed linear and sigmoid activation functions. Among optimizers, RMSprop demonstrated superior stability and performance compared to Adam and Nadam, particularly when paired with ReLU.
Loss Function Performance: The LMSE-based model achieved a higher $R^2$ (0.967) on the testing set, indicating strong alignment with high $N_f$ observations. Conversely, the LMSLE-based model achieved a lower $R^2$ (0.86) but a significantly lower Scatter Index (0.503 for LMSE vs. 0.926 for LMSLE), suggesting that LMSLE better captures the relative order and distribution of samples with lower fatigue life values.
Variable Influence: Partial dependence plots revealed that lower air-void content (ideally below 3–4%) and higher binder content (generally above 5.5%) correlate with extended fatigue life. The models confirmed that high air-void content combined with low binder content markedly reduces fatigue resistance.

Significance and Contributions
The authors claim that this study represents a comprehensive application of neural networks for predicting AC fatigue life in laboratory specimens, distinguished by the open availability of both the code and the dataset via a GitHub repository. Key contributions include:

Data Transparency: Providing an open-access foundation for future research to refine models and potentially expand them to include variables such as asphalt and aggregate types, which were excluded in this study due to their high variability.
Methodological Rigor: Demonstrating the efficacy of optimizing hyperparameters and selecting appropriate loss functions (specifically LMSLE) to handle the logarithmic nature of fatigue data.
Practical Utility: Offering a tool for preliminary estimation of expected fatigue life, enabling engineers to adjust mix design parameters (binder and air-void content) early in the design process to reduce reliance on extensive experimental testing.

Limitations
The paper modestly acknowledges that the dataset is relatively small and imbalanced, with a substantial portion derived from a limited number of studies. This may introduce bias toward the trends of dominant sources and limit generalizability, particularly for extreme $N_f$ values or rare fatigue scenarios. The authors note that the model serves best for mixes resembling the dominant data sources used in training and that further experimental data collection is necessary to enhance model robustness and generalizability.