Thursday 8:30 AM

March 3, 2022

Room: 256A

Location: Anaheim Convention Center

Artificial intelligence (AI) probes data for high-dimensional trends that are hard to identify by conventional analysis. Thus, the key to all AI methods, from natural language processing to computer vision to deep learning, is data. However, for many AI applications, the quantity and quality of data required for optimal outcomes is not understood. One solution is to err on the side of data quantity, amassing large, homogeneous data sets. While this may be viable in the social media realm, it is less feasible for physical science and engineering problems where the data is expensive and often heterogeneous. Fortunately, physical data collected by scientists have several advantages: They are selected for their known relevance to the problem, bounded by a physical basis, expertly acquired, and rich in information. Using examples from microstructural characterization, we will survey the factors that should be considered when designing a materials science data set for AI analysis. We will evaluate the relative importance of data size, data type, and data quality. One encouraging observation is that excellent AI outcomes can often be obtained with surprisingly small data sets.

Machine Learning techniques are utilized to describe material line defects from Laue-spot (µXRD) images. Leveraging a large repository of (3D) discrete dislocation dynamics (DDD) structures (from previous studies), a graph-based auto-encoder is trained to produce feature vectors of single-crystal microstructures. Dislocation structure influence on µXRD imaging is expected to be encoded in these features. µXRD images are simulated using GPU-accelerated ray-tracing techniques which captures Bragg-scattering theory, alloy composition, and expected sensor output. We demonstrate that the evolution of dislocation density with observed XRD peak broadening can be expressed as a function of the DDD feature-space representation. By determining feature importance (feature attribution) within the machine learned model, spreading in Laue-spots is correlated with microstructure. Experimental µXRD images are then used to generate representative microstructures, providing insight into the sample’s defect prevalence. This experimental representation can then be forward simulated by DDD, providing a predictive material response to a given stress.

The ability to efficiently generate microstructure instances corresponding to specified statistical descriptors (two-point statistics) is a crucial component in rigorously studying random heterogeneous materials within the Integrated Computational Materials Engineering and Materials Informatics frameworks. However, the lack of computationally efficient, statistically expressive models for achieving this transformation has severely limited researchers’ ability to construct statistically diverse data-sets for these studies. In this presentation, we present a theoretical and computational framework for generating microstructural instances corresponding to specified two-point statistics by stochastically modeling the microstructure as an N-output Gaussian Random Field. Specifically, we illustrate how two-point statistics can be used to parameterize statistically anisotropic Gaussian Random Fields and we propose the algorithms necessary to efficiently sample these fields (under 0.5 seconds per sample for N phases). Additionally, we address the usefulness of this framework to the future of Materials Informatics.

Currently, the mainstream approach for electron backscatter diffraction (EBSD) pattern simulation is through a physics-based forward model, which first computes the back-scattered yield over all directions, and then generates patterns corresponding to certain orientations through a gnomonic projection. The first stage is time-consuming, limiting its application when there is variation in parameters other than orientation. For discriminative purposes, the EBSD-CNN and EBSDDI-CNN approaches have proved great feature extraction capability of deep neural networks in this domain. Recently, we have shown that a conditional variational autoencoder (CVAE) can realize parametric simulation of EBSD patterns. As a preliminary verification, it takes orientation as the only variable input. In this study, the model is combined with a generative adversarial network (GAN) to realize EBSD pattern simulation over multiple parameters. Compared with the conventional forward model, the deep generative model summarizes the distribution of back-scattered electrons at a higher level.

Prediction of microstructure evolution during material processing is essential to control the material properties. Simulation tools for microstructure evolution prediction based on physical concepts are computationally expensive and time-consuming. Therefore, they are not practical when there is an urgent need for microstructure during the process. Essentially, microstructure evolution prediction is a spatiotemporal sequence prediction problem, where the prediction of material microstructure is difficult due to different process history and chemistry. We propose a generative adversarial networks-long short-term memory (GAN-LSTM) model for the microstructure prediction by combining the generating ability of the GAN with the forecasting ability of the LSTM network. As a case study, we used a dataset from spinodal decomposition simulation of FeCrCo alloy created by the phase-field method for training and predicting the future microstructures by previous observations. The results show that the trained network is capable of efficient prediction of microstructure evolution.

Microstructure evolution-based constitutive models in macro-scale simulation tools require statistical microstructure input at each integration point. This paper reports several robust procedures for interpolating orientation distribution functions (ODFs) from coarsely spaced measurement grids to finely spaced modeling grids. The procedures are based on representing ODFs using generalized spherical harmonics (GSH) functions. Relying on linearity of the representation space, the core of the procedures involves weighting of the GSH expansion coefficients over a given variable such as space, strain, or temperature, thus providing ODFs as a function of location, deformation, or thermal treatment. The procedures are applied to a comprehensive data set obtained by neutron diffraction for a hemispherical part. Utilizing the ODF interpolation procedures, texture is interpolated over the part revealing significant anisotropy in the texture dependent thermal expansion and elastic coefficients. These properties are then used in a thermal finite element-based simulation to predict distortion of the hemispherical part.