AI Digital Twin Detects Hidden Cracks in Composite Materials With 95 Percent Accuracy
A new framework combines finite element simulations and deep learning to monitor structural health and locate damage in aerospace materials.
A new engineering approach combines artificial intelligence and virtual replicas to spot hidden cracks and structural flaws in aerospace composite materials with over 95 percent accuracy. Composite materials are increasingly utilized in demanding engineering fields, notably in aviation, due to their lightweight nature and high strength. However, these structures remain susceptible to hidden problems like internal delamination and cracks during their operational service life, which can compromise structural integrity.
Early and accurate detection of such damage is crucial for ensuring safety and minimizing maintenance costs. Traditional inspection methods can be time-consuming and often fail to catch hidden flaws effectively. A research team featuring scientists Ameni Ragoubi, Tjerk Timan, Ronan Goff, Julien Brocail, and Jean-Baptiste Casimir has developed an artificial intelligence-assisted digital twin framework designed to address this challenge by merging physical sensor data with computer simulations.
How Virtual Simulations Train Artificial Intelligence to Spot Cracks
The newly developed system relies on two validated finite element models created using Abaqus modeling software to simulate physical stress. The first model simulates low-energy impacts on a composite plate measuring 500 millimeters by 200 millimeters with a thickness of 2.5 millimeters, using a hemispherical impactor. The second model captures guided wave propagation across the material using Macro Fiber Composite actuators and sensors.
By running simulations for 2,000 distinct impact scenarios, the researchers generated a comprehensive database linking specific damage shapes with corresponding vibrational responses. These simulated damage scenarios were then categorized into six specific classes. Classes 0 through 3 represent distinct quadrants of the composite plate where damage might occur, Class 4 covers impacts located along the central horizontal and vertical axes, and Class 5 represents a pristine plate with no damage present.
The resulting methodology is detailed in a study published in Advances in Artificial Intelligence in Manufacturing III. The framework allows physical sensors attached to a structure to stream continuous wavelet transform data, which the digital twin immediately compares against the pre-calculated simulation database to identify matching scenarios.
Deep Learning Architecture for Real-Time Structural Health Monitoring
To process the vibrational data and classify the damage, the team trained a Convolutional Neural Network, which is a specialized deep learning architecture designed for pattern recognition. The network structure begins with three one-dimensional convolutional layers that extract local features from time-domain amplitude data. Each convolutional layer is followed by a pooling layer to reduce data dimensionality while retaining essential information and preventing overfitting.
The network utilizes 128 filters in its first convolutional layer, followed by layers with 64 filters, including a dropout rate of 0.3 in the third layer to further enhance generalization. After flattening the extracted features into a one-dimensional vector, the data passes through two fully connected dense layers containing 256 and 128 nodes respectively. A final output layer using a softmax activation function converts these calculations into probabilities across the six distinct damage classes.
When evaluated over 100 training epochs, the precision curve for both training and validation stabilized near 0.95. A confusion matrix analysis revealed a very low error rate across the board. Classes 3, 4, and 5 achieved near-perfect accuracy, while the remaining classes maintained precision scores above 0.9. This uniform performance indicates that the deep learning model treats each classification group fairly without significant bias.
Next Steps for Operational Aerospace Deployment
While the initial results demonstrate high classification accuracy using simulated databases, the authors note several limitations that must be addressed before widespread industrial adoption. The current framework relies on controlled finite-element scenarios and requires further testing under complex, real-world loading conditions that fluctuate dynamically.
Future research will focus on evaluating the computational performance of the framework, specifically measuring inference time and system latency to verify its viability for real-time monitoring in high-stakes operational environments. Additionally, the team aims to refine damage localization by moving beyond categorical quadrants toward regression-based approaches that can predict exact spatial coordinates and precise damage extents.
The research was published in Advances in Artificial Intelligence in Manufacturing III on November 19, 2026.
This article has been fact checked for accuracy, with information verified against reputable sources. Learn more about us and our editorial process.
Last reviewed on .
Article history
- Latest version
Reference(s)
- Ragoubi, Ameni., et al. “AI-Based SHM Method for Composite Structure Based on Simulated Data.” Lecture Notes in Mechanical Engineering, November 19, 2026, pp. 112-119. Springer Nature Switzerland, doi: 10.1007/978-3-032-44811-8_13. <https://doi.org/10.1007/978-3-032-44811-8_13>.
Cite this page:
- Posted by Anjali Singh