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

S. Thompson, F. Teixeira-Dias, M. Paulino, A. Hamilton

Research output: Contribution to journalArticlepeer-review

Abstract / Description of output

Ballistic impacts are a primary risk in both civil and military defence applications, where successfully predicting the dynamic response of a material or structure to impact crucial to the design of safe and fit-for-purpose protective structures. This study proposes a conditional Generative Adversarial Network (cGAN) architecture that can learn directly from available ballistic data and can be conditioned on additional information, such as class labels, to govern its output. A single Multi-Layer Perceptron (MLP) cGAN architecture is trained on a multi-class ballistic training set consisting of 10 classes labelled 0−9 where each class refers to a ballistic curve with a different ballistic limit velocity, vbl. A total of 5 models are trained on datasets consisting of 5, 10, 15, 20 and 25 samples within each class. For integer class labels 0−9, all cGAN models successfully predict the vbl with a maximum error of 4.12%. Additionally, for non-integer class labels between 0−9 the vbl predictions are similar despite not explicitly appearing in the training set. Moreover, each cGAN model is challenged to generate new samples for class labels that exist beyond the scope of the training set for class labels between 9 − 20. Four of the models predict the vbl with an error of less than 1.5% in all cases. This study showcases how a cGAN model can learn directly from a multi-class ballistic dataset and generate additional samples representative of that data for classes that do not appear explicitly in the training set.
Original languageEnglish
Pages (from-to)425-440
JournalNeural Networks
Early online date2 Aug 2022
Publication statusPublished - Oct 2022

Keywords / Materials (for Non-textual outputs)

  • Machine learning
  • Multi-Layer Perceptron (MLP)
  • Conditional Generative Adversarial Networks (cGAN)
  • Terminal ballistics
  • Armour systems
  • Multi-class dataset


Dive into the research topics of 'Predictions on multi-class terminal ballistics datasets using conditional Generative Adversarial Networks'. Together they form a unique fingerprint.

Cite this