Elucidating Grain Growth in Thermo-Magnetic Processed Materials by Transfer Learning and Reinforcement Learning

The goal of the proposed work is to combine deep, model-based reinforcement learning and transfer learning to elucidate one of the most fundamental, yet poorly understood, mechanisms in materials science: abnormal grain growth. We hypothesize that abnormal grain growth is the result of highly anisotropic grain boundary character networks, where a unique combination of neighboring grain boundaries incentivize accelerated growth. We will test this hypothesis with an anisotropic mesoscale grain growth model that incorporates the complexity of grain boundary character and energy anisotropy. Rather than rely on heuristics to define the grain growth behavior, we will teach the mesoscale grain growth model how to simulate grain growth. This is accomplished by integrating experimental microstructure data with model-based machine learning strategies. A machine learning approach is necessary for us to capture the high combinatorial, highly complex space of grain boundary character in a grain growth model. The developed anisotropic mesoscale grain growth model will be a new tool for exploring the microstructural features and processing parameters critical to inducing and sustaining or inhibiting abnormal grain growth. This project addresses two of DOE's main topic areas: (1) synthesis science including nucleation, growth and restructuring of hierarchical materials and (2) behavior of properties and processes in extreme environments, particularly temperature and magnetic fields. This goal will be accomplished by our team's unique expertise, which encompasses grain boundary characterization, machine learning methods, mesoscale grain growth modeling, and thermomagnetic materials processing.

This project will be divided into three tasks using alumina as a model material due to its high anisotropy:

1. Use machine learning methods to identify the physical descriptors of 3D microstructural data collected at different periods of grain growth. The samples will be made with both traditional powder processing and a thermo-magnetic process to create a wide array of grain topologies and crystallographic textures.

2. Apply transfer and reinforcement learning to train the anisotropic mesoscale grain growth model, where the machine learning agents will adjust the simulations to reach the physical descriptors previously identified for each growth period in Task 1.

3. Validate the developed model by conducting a grain growth study using a non-destructive characterization method and directly comparing the results to simulations. We will then test our hypothesis using the validated anisotropic mesoscale grain growth model with simulated microstructures where we control the character anisotropy of neighboring grain boundaries.

The anisotropic mesoscale grain growth model will provide mechanistic insight into grain growth that would otherwise be inaccessible from experiments and isotropic grain growth models. These insights could inspire new processing designs that either promote or avoid these grain boundary character networks to control abnormal grain growth.