AI/Data informatics: Tools for Accelerated Design of High-temperature Alloys: High Temperature Mechanical Properties
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Integrated Computational Materials Engineering Committee, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Michael Titus, Purdue University; Pinar Acar, Virginia Tech; Andrew Detor, DARPA/DSO; James Saal, Citrine Informatics; Dongwon Shin, Oak Ridge National Laboratory
Wednesday 2:00 PM
March 17, 2021
Room: RM 33
Location: TMS2021 Virtual
Session Chair: Andrew Detor, GE Research; Dongwon Shin, Oak Ridge National Laboratory; Sudeepta Mondal, Argonne National Laboratory
2:00 PM
Elastic Properties Machine-learning-based Descriptor for a Refractory High Entropy Alloy: Guillermo Vazquez Tovar1; Prashant Singh2; Daniel Sauceda1; Raymundo Arroyave1; 1Texas A&M University; 2AMES Laboratory
The discovery of new materials is crucial for the development of new and existing technologies, and the infinitely big material design space may hinder better options for modern demand. In this work, we address the need for a computational model for the elastic properties of the alloy system MoNbTaVW. The proposed accurate and computationally inexpensive model is based on using elastic data from density functional theory (DFT) stress-strain calculations to build a machine learning-based descriptor: SISSO (Sure Independence Screening Sparsifying Operator). SISSO does a feature selection of a space made from combinations of atomic features. The final descriptor has an accuracy similar to experimental characterization for the elastic constants C11, and bulk modulus K. Since this method relies on the combination of physical features, the final descriptor returns a physically meaningful expression that contains relevant atomic values to tune for the desired material design.
2:20 PM
Predicting Yield Stress of High Temperature Alloys via Computer Vision and Machine Learning: Nan Gao1; Zongrui Pei2; Youhai Wen2; Michael Gao2; Elizabeth Holm1; 1Carnegie Mellon University; 2National Energy Technology Laboratory
Understanding the linkage between microstructure and properties is especially important to material design for high temperature performance. Generally, microstructures are characterized by visual inspection and metallographic measurements. Although morphology information can be captured and observed, the rich, multiscale microstructural feature data contained in a typical micrograph is rarely fully quantified or exploited. In this research, pre-trained convolutional deep neural networks (CNNs) are used to extract visual information from images, and machine learning methods are trained to make predictions of mechanical properties based on features that exist at a hierarchy of length scales. The temperature-dependent yield stress of steel alloys is predicted with good fidelity, and links to microstructural features that influence mechanical response are made. We find that computer vision and machine learning are promising tools for connecting microstructure to properties.
2:40 PM
Revealing Nanoscale Features Controlling Diffusion Within Multi-component Alloys through Machine Learning: S. Mohadeseh Taheri-Mousavi1; S. Sina Moeini-Ardakani1; Ryan W. Penny1; Ju Li1; A. John Hart1; 1MIT
The immense compositional breadth of non-dilute multi-component and concentrated alloys has made their well-targeted design extremely challenging. Here, we present a newly developed numerical framework whereby deep learning algorithms supervised by atomistic-scale simulations are used to explore the nanoscale features controlling the diffusivity of atomic components in heavily alloyed compounds. Due to inherent non-linear optimization of the machine learning algorithms, the prediction accuracy is at least 10-fold improved over a conventional clustering method. Analysis of all possible atomic configurations and compositions within a model NiAl alloy reveals how the propensity of Al to form short-range-order near vacancies correlates with the generalized stacking fault energy of configurations with mobile Ni atoms. In the future, this approach can guide the selection of composition and processing parameters for conventional as well as additive manufacturing techniques, and it could enable design of metals with tailored gradient diffusivity for high temperature applications.
3:00 PM
Uncertainty Quantification for Thermo-mechanical Behavior of Aircraft Engine Materials in Elevated Temperatures: Arulmurugan Senthilnathan1; Pinar Acar1; 1Virginia Tech
The Turbine inlet temperature (TIT) is a crucial parameter, which affects the overall efficiency of gas turbine engines. However, its indirect measurement leads to uncontrolled uncertainty in thermal and mechanical properties. An important example of turbine materials is the Titanium-Aluminum alloys owing to their resistance to high thermal and mechanical stresses. Consequently, the goal of the present study is to understand the changes in microstructural and thermo-mechanical behavior of Titanium-Aluminum alloys at TIT, by considering the effects of the uncertainty. The mechanical response of the alloy is computed using LAMMPS for a range of elevated temperature values to incorporate the uncertainty in TIT. The propagation of the TIT uncertainty on the mechanical properties is identified with the computed range for the stress-strain response. The future work will utilize these physics-based simulations to develop a machine learning model that can predict the stress-strain behavior of the alloy at a given temperature.
3:20 PM Invited
Coupling of Data Mining, Thermodynamics and Multi-objective Genetic Algorithms for the Design of High-temperature Alloys: Franck Tancret1; Edern Menou2; Gérard Ramstein1; 1University of Nantes; 2Safran
Our use of artificial intelligence (AI) to design high-temperature alloys (HTA) started more than twenty years ago with the modelling of mechanical properties of nickel-based superalloys as a function of composition, using data mining tools like artificial neural networks and Gaussian processes (GP). Such machine learning (ML) models were then associated to the calculation of phase diagrams (Calphad / Thermo-Calc) to design an affordable wrought superalloy for power plant applications, and later a set of single-crystal superalloys for aeroengines. Other AI tools, like genetic algorithms (GA), including their multi-objective optimisation (MOO) version, were then coupled to both ML and Calphad to propose the most advanced integrated computational alloy design scheme at its time, along with the successful redesign of a proprietary superalloy for turbine disks. Current works are ongoing, both on algorithm development and on the design of HTAs like so-called high entropy alloys (HEA) or complex concentrated alloys (CCA).
3:50 PM
Machine Learning Augmented Predictive & Generative Models for Rupture Life in High Temperature Alloys: Madison Wenzlick1; Osman Mamun2; Ram Devanathan2; Kelly Rose3; Jeffrey Hawk3; 1Leidos Research Support Team for the National Energy Technology Laboratory; 2Pacific Northwest National Laboratory; 3National Energy Technology Laboratory
Exploring the connections between material pedigree and performance is critical to understanding creep behavior. This work leverages the data framework for collection, curation and processing of alloy data established through DOE’s eXtremeMAT project. This work investigates both the semi-empirical and data-driven methods of predicting rupture life. Gradient boosting machine learning algorithms are applied to predict rupture time by first predicting the Larson-Miller Parameter, a commonly applied metric for evaluating creep behavior, as well as directly modeling rupture time. The models were evaluated using high quality 9-12% Cr ferritic-martensitic steel data and the most effective model was applied to austenitic stainless steels. A generative model was applied to generate synthetic data within the alloy space to evaluate the effectiveness of supplementing the dataset with synthetic information. A workflow for incorporating data generation for alloy design is described.
4:10 PM
Determining Solute Site Preference and Correlations to Antiphase Boundary Energy in Ni-based Superalloys: Enze Chen1; Tao Wang2; Mario Epler2; Timofey Frolov3; Mark Asta1; 1University of California, Berkeley; 2Carpenter Technology Corporation; 3Lawrence Livermore National Laboratory
Ni-based superalloys are a superior class of structural materials used in aircraft turbines and power plants due to their excellent strength, creep resistance, and corrosion resistance at high temperatures. In particular, their high-temperature strength is linked to high antiphase boundary (APB) energy in the Ni3Al precipitates, which motivates a better understanding for how chemical heterogeneity affects the APB energy. The APB energy varies with not only solute chemistry and concentration, but also sublattice site preference in the ordered (L12 structure) Ni3Al precipitates. We use a thermodynamic model implemented in PyDII combined with density functional theory calculations to predict the site preference of alloying additions to Ni3Al and derive descriptors that correlate with high APB energy. We discuss how this methodology allows us to intelligently screen for promising superalloy chemistries through validation with a subset of common alloying elements.