A joint voxel flow - phase field framework for ultra-long microstructure evolution prediction with physical regularization

Abstract: Phase-field (PF) modeling is a powerful tool for simulating microstructure evolution. To overcome the high computational cost of PF in solving complex PDEs, machine learning methods such as PINNs, convLSTM have been used to predict PF evolution. However, current methods still face shortages of low flexibility, poor generalization and short predicting time length. In this work, we present a joint framework coupling voxel-flow network (VFN) with PF simulations in an alternating manner for long-horizon temporal prediction of microstructure evolution. The VFN iteratively predicts future evolution by learning the flow of pixels from past snapshots, with periodic boundaries preserved in the process. Periodical PF simulations suppresses nonphysical artifacts, reduces accumulated error, and extends reliable prediction time length. The VFN is about 1,000 times faster than PF simulation on GPU. In validation using grain growth and spinodal decomposition, MSE and SSIM remain 6.76% and 0.911 when predicted 18 frames from only 2 input frames, outperforming similar predicting methods. For an ultra-long grain growth prediction for 82 frames from 2 input frames, grain number decreases from 600 to 29 with NMSE of average grain area remaining 1.64%. This joint framework enables rapid, generalized, flexible and physically consistent microstructure forecasting from image-based data for ultra-long time scales.