MICROSTRUCTURE EVOLUTION PREDICTION USING DEEP LEARNING

Abstract

Video prediction models can be used alongside phase field simulations to mitigate some of inherent numerical constraints. A modified version of a simpler yet better video prediction (SimVP) model was used to predict microstructure evolution based on phase field simulations. The model showed good performance that aligned with the ground truth to a great extent in terms of pixel-wise quantitative metrics (mean squared error and mean absolute error), perceptual metrics (structural similarity index metric and peak signal to noise ratio) and physics based metrics (energy, number of grains, grain radius, etc.). We found that the modified SimVP model outperformed a comprobable alternative model, the E3D model, for spinodal decomposition and dendritic growth with a significant margin. The model produced realistic looking microstructure for the three processes even in the long term.