Application of generative-adversarial networks in modeling plasma-chemical processes for obtaining nanosystems

The use of neural networks makes it possible to take into account the nonlinearity and complex relationships between the parameters of the technological process, which is multifactorial, with dependencies between the parameters that are difficult to formalize by classical methods. To generate additional data in modeling physiochemical processes for the production of nanosystems, it is proposed that generative adversarial networks (GANs) be used. The problematic aspects of the process of obtaining such training arrays are considered, in particular: the quality of synthetic data, preservation of functional dependencies between input and output data, and stable training of the generator and discriminator. The settings of neural network training parameters are analyzed: the value of noise in the generator, the size of the latent space, activation functions, and tracking the boundaries of synthetic data after generation. The necessity of adapting neural network approaches to the specifics of the process is substantiated. The search for the optimal architecture of the generative-adversarial network and parameter settings are considered. For example, a data augmentation strategy is required to train no-deep neural networks by generating synthetic data. It is proved that the lack of sufficient experimental data due to the high cost and complexity of conducting experiments can be compensated by fake generated data. In this paper, we design and implement an appropriate GAN architecture. The resulting synthetic sample is used to train no-deep neural networks, and the real data set is used to test. The expediency of using generative-adversarial networks to obtain synthetic data, which were subsequently used to train a no-deep neural network for modeling the parameters of plasma chemical processes of nanoparticle production, was demonstrated.