Inferring particle distributions in two-dimensional space with numerical features based on generative adversarial networks (GANs)

A feasibility study was conducted on the usage of Generative Adversarial Networks (GANs) for inferring particle beam profiles. Two types of GANs, Deep Convolution GAN (DCGAN) and Wasserstein GAN (WGAN), were implemented in the PyTorch framework and trained using a mathematically generated dataset. The input latent vector represents an ensemble of features that defines unique probabilities, indicating the degrees to which the data belongs to specific categories. It was shown that the GANs are able to reproduce successfully with the given features having ± 20 % uncertainty. The same architectures for the generator and discriminator showed different performances depending on the learning schemes in the performance evaluations; DCGAN showed smaller error fluctuations compared to WGAN. Meanwhile, WGAN generated better images for the convolution of two distributions provided with the pairs of corresponding latent vectors, whereas DCGAN produced artificial anomalies in its results. This implies that WGAN strengthens the robustness of the generator. The GANs demonstrated its functionality as a regression model for unidentified distributions, highlighting the potential applications of generative networks in analyzing complex and irregular behaviors of particle beams in related fields.