Efficient prediction of phase-field crystal dynamics via β-variational autoencoders and time-series transformers on coupled physical fields

초록

Dendritic crystal growth is a complex phenomenon that has traditionally required high-fidelity simulations, which are computationally expensive. This study introduces a data-driven reduced order modeling framework for efficient prediction of dendritic crystal growth. A beta-variational autoencoder is utilized to compress coupled physical fields into a compact latent space. We systematically evaluate the impact of the regularization parameter and latent dimensionality on reconstruction accuracy. The trained encoder-decoder pair is integrated into an end-to-end time-series forecasting framework, where multiple representative models are employed to predict future latent dynamics. We investigate the influence of input sequence length and prediction horizon on forecasting accuracy, as well as the inference efficiency of the different models. Numerical experiments on a phase-field crystal growth dataset demonstrate that the proposed approach achieves high reconstruction fidelity, robust predictive performance, and significant reduction in computational cost. This offers a practical solution for fast modeling and multi-scale dynamics prediction in complex physical systems.

키워드