A New Method to Predict Damage to Composite Structures Using Convolutional Neural Networks

Area of Science:

• Materials Science
• Mechanical Engineering
• Computational Science

Background:

• Virtual testing and finite element methods (FEMs) are crucial for predicting damage in composite aeronautical structures.
• FEMs for composite impact analysis are computationally intensive and require significant expertise.
• Artificial intelligence offers a promising alternative for efficient composite damage assessment.

Purpose of the Study:

• To develop and evaluate a deep learning methodology for predicting impact damage in fiber-reinforced polymer (FRP) composites.
• To explore the effectiveness of convolutional neural network (CNN) architectures for this prediction task.
• To demonstrate the potential of AI in accelerating the design process for composite structures.

Main Methods:

• A deep learning approach utilizing CNN models was developed.
• Data was sourced from literature and FEM simulations, augmented to increase dataset size from 149 to 2725 instances.
• Two CNN architectures were aggregated and compared against a single CNN model.

Main Results:

• The aggregated CNN model outperformed a single CNN architecture.
• The model achieved a precision of 0.15 mm for length measurements.
• Average delaminated surface error was 56 mm², with a 7% error rate for delamination presence prediction.

Conclusions:

• Aggregated CNN models show strong potential for accurately predicting impact damage in composites.
• This AI-driven approach can significantly accelerate the design cycle for aeronautical composite structures.
• The methodology offers a more efficient alternative to traditional FEM-based virtual testing.