Machine Learning Based Methods for Obtaining Correlations between Microstructures and Thermal Stresses

The microstructure–property relationship is critical for parts made using the emerging additive manufacturing process where highly localized cooling rates bestow spatially varying microstructures in the material. Typically, large temperature gradients during the build stage are known to result in significant thermally induced residual stresses in parts made using the process. Such stresses are influenced by the underlying local microstructures. Given the extensive range of variations in microstructures, it is useful to have an efficient method that can detect and quantify cause and effect. In this work, an efficient workflow within the machine learning (ML) framework for establishing microstructure–thermal stress correlations is presented. While synthetic microstructures and simulated properties were used for demonstration, the methodology may equally be applied to actual microstructures and associated measured properties. The dataset for ML consisted of images of synthetic microstructures along with thermal stress tensor fields simulated using a finite element (FE) model. The FE model considered various grain morphologies, crystallographic orientations, anisotropic elasticity and anisotropic thermal expansion. The overall workflow was divided into two parts. In the first part, image classification and clustering were performed for a sanity test of data. Accuracies of 97.33% and 99.83% were achieved using the ML based method of classification and clustering, respectively. In the second part of the work, convolution neural network model (CNN) was used to correlate the microstructures against various components and measures of stress. The target vectors of stresses consisted of individual components of stress tensor, principal stresses and hydrostatic stress. The model was able to show a consistent correlation between various morphologies and components of thermal stress. The overall predictions by the model for all the microstructures resulted into R2≈0.96 for all the stresses. Such a correlation may be used for finding a range of microstructures associated with lower amounts of thermally induced stresses. This would allow the choice of suitable process parameters that can ensure that the desired microstructures are obtained, provided the relationship between those parameters and microstructures are also known.

Microstructure Comparison Using Dissimilarity Metric for Additive Manufacturing of Metals

Additive Manufacturing (AM) simulations are often employed to replace the expensive experiments to study the effects of processing conditions. In process modeling, one of the key limitations is the lack of reliable validation techniques. The stochastic nature and the spatial heterogeneity of microstructures make it difficult to validate the simulated microstructures against experimentally obtained images through statistical measures (e.g. average and standard deviation of grain sizes). In this work, a validation metric is proposed that can effectively quantify the dissimilarity between two AM microstructures. The methodology involves first calculating the Angularly Resolved Chord Length Distribution (ARCLD) at representative angles and then computing the Earth Mover’s Distance (EMD) to obtain the final unitless score that is named Dissimilarity Score (DS). The efficacy of the proposed methodology was first tested on synthetic microstructures, and then on AM simulations that employ the solidification model-Cellular Automaton (CA) with IN625. Results show that DS effectively measures the dissimilarity between different microstructures. The use of DS is also extended to calibrate the CA processing simulation code to match with experimental AM images from NIST AM-Bench Challenge.

A Novel Approach to Discrete Representative Volume Element Automation and Generation-DRAGen

In this study, a novel approach for generating Representative Volume Elements (RVEs) is introduced. In contrast to common generators, the new RVE generator is based on discrete methods to reconstruct synthetic microstructures, using simple methods and a modular structure. The plain and uncomplicated structure of the generator makes the extension with new features quite simple. It is discussed why certain features are essential for microstructural simulations. The discrete methods are implemented into a python tool. A Random Sequential Addition (RSA)-Algorithm for discrete volumes is developed and the tessellation is realized with a discrete tessellation function. The results show that the generator can successfully reconstruct realistic microstructures with elongated grains and martensite bands from given input data sets.

Prediction of Cyclic Stress–Strain Property of Steels by Crystal Plasticity Simulations and Machine Learning

In this study, a method for the prediction of cyclic stress–strain properties of ferrite-pearlite steels was proposed. At first, synthetic microstructures were generated based on an anisotropic tessellation from the results of electron backscatter diffraction (EBSD) analyses. Low-cycle fatigue experiments under strain-controlled conditions were conducted in order to calibrate material property parameters for both an anisotropic crystal plasticity and an isotropic J2 model. Numerical finite element simulations were conducted using these synthetic microstructures and material properties based on experimental results, and cyclic stress-strain properties were calculated. Then, two-point correlations of synthetic microstructures were calculated to quantify the microstructures. The microstructure-property dataset was obtained by associating a two-point correlation and calculated cyclic stress-strain property. Machine learning, such as a linear regression model and neural network, was conducted using the dataset. Finally, cyclic stress-strain properties were predicted from the result of EBSD analysis using the obtained machine learning model and were compared with the results of the low-cycle fatigue experiments.