Generative artificial intelligence in smart manufacturing

process continuous, performance of the two neural networks improves and the generated output become closer to the desired one.GANs suffer from the model collapse manifesting itself by the generator model not fully capturing the diversity of the data.This weakness has been addressed with extensions such as a multi-agent GAN involving multiple generators and one discriminator.Variational autoencoders (VAEs) were developed in 2014, the same year as GANs (Kingma & Welling, 2022).A VAE model is based on the concept of variational inference (approximation of complex distributions) and it generates new data that resembles the training data.A VAE includes two neural networks: an encoder and a decoder.The encoder transforms the input, while the decoder reconstructs the input from the transformed data.VAEs are used in large language models (LLMs) to generate text and in applications ranging from image generation to anomaly detection.Diffusion models are largely applied to generate optimized images.The generated data has a more complex and meaningful data distribution than the original one.A flow-based model transforms data from a simple base distribution (e.g., Gaussian) into the target distribution (e.g., a pattern).The invertible transformation is referred to as flow.Computational efficiency and generalization capability make them suitable for applications in computer vision, language processing, anomaly detection, and generative design.Transformer neural network models (referred to as transformers) are used in signal processing, natural language processing, computer vision, audio and speech processing, as well as multi-modal tasks (Islam et al., 2024).Transformers were introduced by Vaswani et al. (2017) based on the previously developed attention mechanism capturing the contextual relationships in sequential data.