Georgia Institute of Technology Materials Science and Engineering

Committee Members: 

Prof. Surya Kalidindi, Advisor, MSE

Prof. Matthew McDowell, MSE

Prof. Hamid Garmestani, MSE

Prof. Roshan Joseph, ISyE

Balasubramaniam Radhakrishnan, Ph.D., ORNL

Jim Belak, Ph.D., LLNL

“Gaussian Process-based Workflows for Efficient and Autonomous Extraction of Process-structure-property Linkages”   

Abstract:

Materials processing consists of a vast range of operating conditions, which can have a profound impact on the resulting material structure. In addition, the range of possible microstructural features and morphologies is also extremely large. Therefore, an approach is required to explore these spaces and produce materials knowledge in an efficient manner. This work seeks to address this problem with a new process-structure-property (PSP) framework based on Gaussian Process Regression (GPR). GPR is able to produce predictions with uncertainty quantification, which is utilized for the adaptive selection of new data. By implementing the GPR-driven framework on a high-performance computing cluster, materials knowledge is generated on a full-time basis, in a completely autonomous manner.

The first project demonstrates a novel strategy for the autonomous development of a machine-learning model for predicting the equivalent stress–equivalent plastic strain response of a two-phase composite calibrated to micromechanical finite element models. The model takes a user-defined three-dimensional, two-phase microstructure along with user-defined hardening laws for each constituent phase, and outputs the equivalent stress–plastic strain response of the microstructure modeled using J2-based isotropic plasticity theory for each constituent phase. It is demonstrated that the use of GPR together with a Bayesian sequential design of experiments can lead to autonomous protocols for optimal generation of the training dataset and the development of the model. It is shown that this strategy dramatically reduces the time and effort expended in generating the training set.

The second project seeks to extend the framework of the first project to process-structure linkages, which involve the time evolution of multiphase microstructures. Gaussian Process Autoregression (GPAR) is used to predict solid-state phase transformations of Ni-based superalloys in the context of an additive manufacturing process. These alloys contain multiple inclusion phases, with each inclusion phase having multiple crystallographic variants. It is shown that, even for a very large number of local states, reduced-order microstructure representations show clear differences based on their corresponding processing parameter values. This result is promising for the development of suitable GPAR models.

The third project looks to implement an exascale computing approach for the development of GPAR models. Currently, there are two major computational bottlenecks in the construction of the models.

Principal component analysis (PCA), which is used for the development of reduced-order microstructure representations, becomes very computationally expensive as the dimensionality of the dataset increases. Training GPAR models, which requires the inversion of a covariance matrix, becomes expensive as the number of microstructures in the dataset increases. Using parallelized, GPU-accelerated approaches to address these issues provides the possibility for developing large-scale materials knowledge workflows