Combining phase field modeling and deep learning for accurate modeling of grain structure in solidification

Abstract

Additive manufacturing by wire deposition is a complex process as it generates overlapping and transient thermal fields, resulting in multiple cycles of solidification and remelting. Consequently, simulating the microstructure becomes challenging, making it difficult to optimize the process and predict the mechanical properties of an additive-manufactured part. In this work, a framework is developed, including several sequentially applied methods to obtain increasingly accurate information about the microstructure resulting from a wire deposition process. This framework is based on understanding experimental data and utilizing them as inputs to simplify the problem or accelerate numerical simulations. The experimental inputs consist of light optical images, SEM, and EBSD images. The methods developed here are initially geometric-based and then progress to shape recognition, unsupervised learning, and phase field modeling.