Predictions on multi-class terminal ballistics datasets using conditional Generative Adversarial Networks.

Area of Science:

• Computational Mechanics
• Materials Science
• Artificial Intelligence

Background:

• Predicting material and structural response to ballistic impacts is critical for defense and civil applications.
• Existing methods may struggle with the dynamic and complex nature of ballistic data.
• Generative Adversarial Networks (GANs) offer potential for learning complex data distributions.

Purpose of the Study:

• To propose and evaluate a conditional Generative Adversarial Network (cGAN) for ballistic impact analysis.
• To assess the cGAN's ability to predict ballistic limit velocity (vbl) from limited data.
• To investigate the cGAN's capacity for generating new ballistic data samples for unseen classes.

Main Methods:

• A Multi-Layer Perceptron (MLP) based cGAN architecture was developed.
• The cGAN was trained on a multi-class ballistic dataset with 10 labeled classes (0-9).
• Models were trained using varying dataset sizes (5 to 25 samples per class).

Main Results:

• cGAN models accurately predicted vbl for integer class labels (0-9) with a maximum error of 4.12%.
• Predictions for non-integer class labels (0-9) were accurate despite not being in the training set.
• cGANs generated new samples for class labels beyond the training scope (9-20), with 4 models achieving <1.5% error.

Conclusions:

• The proposed cGAN effectively learns from multi-class ballistic data.
• The cGAN can generate representative ballistic data samples for classes not explicitly present in the training set.
• This approach enhances the predictive modeling of ballistic impacts and supports the design of protective structures.