Using a Convolutional Neural Network to Identify the Microstructural Features that Lead to Anisomorphic Grain Growth

Grain growth, a common mechanism in ceramic processing, is difficult to predict because it is influenced by many factors, including porosity, raw powder size and shape, impurities and sintering conditions. To parse out these effects, grain growth studies commonly rely on qualitative comparisons of micrographs because of inadequate physical descriptors of grain size and shape. However, new data techniques can extract high order parameters that better describe microstructural features. In this study, we establish whether a convolutional neural network (CNN) can distinguish two initial microstructures of calcia-doped alumina, which appear similar by traditional grain growth metrics but then evolve differently during grain growth. If a unique signature can be extracted, it can provide mechanistic-insight into grain boundary motion for improved grain growth predictions.<br/>The goal here is to collect a signature that distinguishes two alumina microstructures prepared with the same raw materials, slip casting process, and sintering conditions. However, one sample is slip cast in a magnetic field, which causes the equiaxed particles to rotate and<br/>preferentially align the c-axis of their unit cell, resulting in a bulk crystallographic texture. Recent studies show that the textured microstructure evolves differently from the untextured sample, resulting in a larger grain size and unique grain morphology. Here, we will evaluate the<br/>developed CNN used to identify these microstructures. Briefly, the CNN is trained from at least five micrographs with ~200 grains each from each timestep. The CNN consists of two convolutional layers with an increasing number of filters to learn increasingly abstract features.<br/>The convolutional layers are followed by three fully connected layers with a decreasing number of nodes. The signature is extracted as a scalar quantity from the penultimate layer in the network. We will discuss the accuracy of the CNN to predict the microstructure’s processing<br/>path and its implications regarding how crystallographic orientation influences grain growth. Furthermore, we will demonstrate the need for a new higher order physical descriptor of microstructures.