Algorithm Development in Materials Science and Engineering: AI/ML Algorithms and Applications
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS Functional Materials Division, TMS Structural Materials Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Solidification Committee, TMS: Chemistry and Physics of Materials Committee
Program Organizers: Adrian Sabau, Oak Ridge National Laboratory; Ebrahim Asadi, University of Memphis; Enrique Martinez Saez, Clemson University; Garritt Tucker, Colorado School of Mines; Hojun Lim, Sandia National Laboratories; Vimal Ramanuj, Oak Ridge National Laboratory
Thursday 8:30 AM
March 23, 2023
Room: Cobalt 502B
Location: Hilton
Session Chair: Enrique Saez, Clemson University; Vimal Ramanuj, Oak Ridge National Laboratory
8:30 AM
Applications of Min-cut Algorithms for Image Segmentation and Microstructure Reconstruction: Austin Gerlt1; Alexander Brust2; Eric Payton3; Stephen Niezgoda1; 1The Ohio State University; 2DNV; 3Air Force Research Lab
Graph cut algorithms are a useful set of techniques for separating out or clustering connected data. In particular, the Boykov Min-cut/Max-flow algorithm is adopted to efficiently run on uniform 2d or 3d grids of data, such as would be seen in an image or 3d raster dataset. These techniques see significant usage in both the medical imaging community (for example, identifying tumors in MRIs) and in ML image processing (for example, object detection). This talk will show several cases in Material Science fields wherein this algorithm can be used to efficiently segment datasets that other algorithms such as watershed, edge detection, or basic thresholding might fail. Notably, this is used for automated spot detection in far field HEDM, and prior Austenite parent reconstruction on Martensitic EBSD scans.
8:50 AM
Machine Learning Models of Effective Properties with Reduced Requirements on Microstructure: Marat Latypov1; 1University of Arizona
Microstructure--property relationships are key to effective design of structural materials for advanced applications. Advances in computational methods enabled modeling microstructure-sensitive properties using 3D models (e.g., finite elements) based on microstructure representative volumes. The need in high-resolution 3D microstructure data limits a wider adoption of microstructure-sensitive 3D models. In this work, we present machine learning (ML) strategies that are less demanding in terms of 3D microstructure input. We will first discuss ML approaches to modeling effective properties of two-phase materials directly from 2D microstructure sections. We then present ML models for mechanical properties of polycrystalline materials based on graph representations of polycrystals.
9:10 AM
Microstructure-Sensitive Calculations of Metal Nanocomposite Electrical Conductivity: William Frazier1; Bharat Gwalani1; Julian Escobar Atehortua1; Joshua Silverstein1; Keerti Kappagantula1; 1Pacific Northwest National Laboratory
Recent literature on copper/graphene composites indicates that graphene additions can improve the conductive properties of polycrystalline copper. To ascertain the significance of these additions and disambiguate their contribution from microstructure on net conductivity, we have developed a finite difference-based calculation of electrical conductivity as a function of microstructure. The effects of grain size, texture, dislocation density, graphene distribution within the bulk, and graphene-grain boundary coverage on the conductivity of copper/graphene nanocomposites were evaluated. A parameter study evaluated the effect of graphene on the copper grain boundaries and within bulk copper on the associated change in net composite electrical conductivity. Calculations indicate that high grain boundary coverage with graphene at grain sizes smaller than 1 μm could appreciably increase the composite bulk electrical conductivity. Twinned grain boundaries slightly improved net electrical conductivity for an equivalent average grain diameter. The effect of intragranular graphene additions are also discussed.
9:30 AM
Persistent Homology for Topological Quantification of Microstructure: Simon Mason1; Dennis Dimiduk2; Steve Niezgoda1; 1Ohio State University; 2BlueQuartz Software LLC
Given the central role of microstructure in property and processing relationships in materials science, having quantitative metrics for the description and characterization of microstructure is essential for the continued development of microstructure-sensitive design. Traditional statistical microstructure descriptors (SMDs), such as n-point statistics, allow for the analysis of spatial relationships of features, capture key short range geometrical features but are largely insensitive to long range connectivity . Topological descriptors built on persistent homology capture this connectivity by tracking the appearance and disappearance of connected components and holes in multiple dimensions as persistence diagrams, and related reduced order metrics such as persistence landscapes. Persistence landscapes in particular are well suited for formulating statistical tests for comparing microstructure. These topological measures can also serve as quality metrics for microstructure generation, either through traditional methods such as using DREAM.3D or to assess the development of novel ML methods.
9:50 AM Break
10:10 AM
Thermographic Process Classification in Electron Beam Additive Manufacturing via Stacked Long Short-Term Memory Networks: Benjamin Stump1; Alex Plotkowski1; Vincent Paquit1; 1Oak Ridge National Laboratory
Additive manufacturing (AM) provides opportunities to produce complex geometries and high-performance materials with an unprecedented amount of control. AM simulations must either choose accuracy or performance; therefore, collecting and analyzing in situ process data offers a tractable way to correlate the process to the results. Previous work successfully correlated noisy, low framerate data to the process classification for a single layer; however, this approach broke down when applying it to other layers potentially due to overfitting the solutions. This work utilized a machine learning approach known as long-short term memory networks (LSTMs) to the same problem. LSTMs, which are known for their ability to deal with time series data, achieved superior results with no parameter turning with the results transferring well to layers it was not trained on. Finally, stacked LSTMs, a technique used in natural language processing, achieve the best results with a lowed bound classification accuracy of 96%.
10:30 AM
Prediction of Cutting Surface Parameters in Punching Processes aided by Machine Learning: Adrian Schenek1; Marcel Görz1; Mathias Liewald1; 1Institute for Metal Forming Technology
Punching represents one of the most frequently used manufacturing processes in the sheet metal processing industry. As an important quality criterion for shear cutting processes, the geometric shape of the cutting surface is considered. In this regard, the edge draw-in height, the clean cut proportion, the fracture surface height and the burr are relevant parameters for monitoring the production quality in punching processes. These parameters can easily be measured in shear cutting processes with an open cutting line (e.g. using laser triangulation). For processes with a closed cutting line, however, such a measurement is often not possible due to the limited accessibility. The present paper therefore proposes a machine learning approach, which enables a data-driven prediction of cutting surface parameters based on measurable process data. The new approach presented in this paper is to pre-train a neural network on numerically determined cutting force curves. As an output, the neural network predicts the mentioned quality parameters of punched sheet metal component edges. The output of the numerically pre-trained neural network is evaluated for numerically and experimentally determined process data and cutting surface parameters.