VAE-ALSTM for prediction of high-dimensional physical field evolution in complex manufacturing: a case study of spinning process

In complex manufacturing processes, the prediction of physical field evolution is critical to precise process control and product quality assurance. To this end, this study proposes a spatiotemporally collaborative prediction approach based on spatial encoding and temporal modeling. Guided by this approach, a hybrid network integrating Variational Autoencoder and Attention-based LSTM (VAE-ALSTM) is constructed. The network maps high-dimensional physical field variables to a compact latent manifold space through a Variational Autoencoder (VAE). In the latent manifold space, an integrated MLP-LSTM-Attention module (ALSTM) is used to simultaneously capture temporal irreversible accumulation processes and external environmental disturbances, thereby forming an end-to-end prediction model suitable for physical field evolution in complex manufacturing processes. To validate the approach, the deformation field evolution of the flange in spinning forming is used as a representative case study. Through comparation with the informed prediction methods in the spinning domain, it is found that VAE-ALSTM can effectively maintain the temporal dependency characteristics of physical field evolution while preserving complete spatial topological information. Overall, the results suggest that VAE-ALSTM provides a promising modeling paradigm for high-dimensional spatiotemporal physical-field prediction, with its effectiveness demonstrated on a representative spinning case.