Spatiotemporal microstructure evolution during martensitic transformation in titanium alloys using deep learning

We address the issue of accuracy and efficiency (speed) in phase field simulations of titanium alloys using heterogeneous microstructures via deep learning based surrogate models. A data set with 124 groups of 3D images ($\sim$320,000 2D images) is first assembled via high throughput phase field simulations by varying the interfacial mobility and interfacial energy. The data is used to train surrogate models to “learn” the spatiotemporal evolution of microstructures. These models predict images over a wide time span by learning from the same number of images from the previous time interval. This also holds when the model learns from images obtained using a low interfacial mobility parameter to predict images with high mobility. Moreover, compared to the typical long short-term memory neural network designed for sequential data, the proposed model shows advantages in both accuracy and efficiency, in predictions of images far from those used in training. Specifically, for predicting the image at 4000th evolved time step, the mean squared error based on pixel value is reduced from 0.2755 to 0.065 (a 76.4% reduction) while the prediction time required is only 1/15, i.e., reduced from 5.11 s to 0.38 s. The work sheds light on the use of deep learning tools to accelerate materials simulations without sacrificing accuracy.