Microstructural investigation of Ti-6Al-4V using phase field modeling and image-driven machine learning

Baskaran, Arun
Thumbnail Image
Other Contributors
Lewis, Daniel
Hull, Robert, 1959-
Maniatty, Antoinette M.
Wen, John T.
Sundararaman, Ravishankar
Issue Date
Materials engineering
Terms of Use
Attribution-NonCommercial-NoDerivs 3.0 United States
This electronic version is a licensed copy owned by Rensselaer Polytechnic Institute, Troy, NY. Copyright of original work retained by author.
Full Citation
Ti64 exhibits a wide variety of morphologies, such as lamellar, bi-modal, martensitic, etc. Thus, a quantitative description of the Ti64 microstructure to augment the qualitative assessment by the human expert is a challenging task. This thesis details a research effort that has focused on augmenting the human expert’s inference of microstructure images through the implementation of an image-driven machine learning task. The overall goal of the effort is to perform contextual feature segmentation of microstructure images. The first part of the effort was coarse-grained classification of input images into one of three microstructure classes, for which a convolutional neural network (CNN) was trained. The second part of the effort was the automated selection of feature segmentation algorithm, such as Histogram of Oriented Gradients (HOG) or marker-based watershed, based on the classification label assigned to each image after the first step. Repeated trials of training the model and testing on an external dataset provided an average classification accuracy of 93%. The work detailed in this thesis is a reflection of the convergence of physics-based simulations and artificial intelligence towards advancements in microstructure science.
Microstructure control is an important part of the material design strategy, and is a central component in the process-structure-property framework. The response of a material’s properties to a particular processing condition can be rationalized through the influence of the processing condition on the microstructure. Thus, an iterative design process towards a target set of properties can be made more efficient through robust microstructure analysis. This thesis details the convergent research output from multi-phase field modeling and image driven machine learning towards understanding the role of microstructure in the process-structure-property linkage in Ti-6wt%Al-4wt%V (Ti64). Whereas the first research effort models the influence of thermal processing on microstructure evolution in Ti64, the second research effort develops a method to extract relevant quantitative information from a given microstructure image.Ti64 is a dual phase alloy at room temperature, with the thermodynamically stable phases being α (hexagonal close packed) and β (body centered cubic).
A multi-component, multi-phase field model is implemented to study the influence of continuous cooling on the solid state phase transformation from β → α, occurring in Ti64. The phase field model is comprised of non-conserved order parameters, each of which uniquely describe an orientational variant of α, and conserved order parameters, each of which describe the composition fields. The model is composed of four parts, each of which address a specific part of the phase transformation. First, the model is coupled to real thermodynamic databases, which facilitate a realistic solute partitioning between the phases in the simulations. Second, the microelasticity theory is incorporated into the model which enables the calculation of the elastic strain interaction between α and β, and between different variants of α. Third, the experimentally observed habit-planes for each variant are incorporated into the model, which preserves the orientation relationship between the variants in the simulations. Fourth, nucleation of α phase from the β matrix is simulated by explicitly adding nuclei at a rate consistent with the classical nucleation theory. The free energy of nucleation is composed of both the chemical free energy difference between the two phases and the local elastic interaction energy. Four different simulations, using this framework, have been detailed in this work. A case study for integration of experimental data and simulation data was demonstrated by fitting the growth rate of α-lamellae calculated from simulations to the corresponding values obtained from in-situ experiments. The fitting study as conducted for different cooling conditions. The effect of two distinct processing variables, cooling rate and prior-β grain size, on the variant distribution was studied. Repeated simulation trials show that the prior-β grain size has a greater effect on the variant distribution than the cooling rate.
December 2020
School of Engineering
Dept. of Materials Science and Engineering
Rensselaer Polytechnic Institute, Troy, NY
Rensselaer Theses and Dissertations Online Collection
CC BY-NC-ND. Users may download and share copies with attribution in accordance with a Creative Commons Attribution-Noncommercial-No Derivative Works 3.0 License. No commercial use or derivatives are permitted without the explicit approval of the author.