Session Sheet - Uncertainty Quantification in Data-Driven Materials and Process Design: Materials Design under Uncertainty

Uncertainty Quantification in Data-Driven Materials and Process Design: Materials Design under Uncertainty
Sponsored by: TMS: Integrated Computational Materials Engineering Committee
Program Organizers: Yan Wang, Georgia Institute of Technology; Raymundo Arroyave, Texas A&M University; Anh Tran, Sandia National Laboratories; Dehao Liu, Binghamton University

Monday 2:00 PM
October 10, 2022
Room: 310
Location: David L. Lawrence Convention Center

Session Chair: Edwin Garcia, Purdue University; Wei Xiong, University of Pittsburgh; Aaron Tallman, Florida International University

2:00 PM
Machine Learning of Phase Diagrams: Jarrod Lund¹; Haoyue Wang¹; Edwin Garcia¹; Richard Braatz²; ¹Purdue University; ²MIT
    By starting from experimental- and ab initio-determined phase diagrams (PDs) of materials, a machine learning (ML) method is developed to infer the free energy function for each phase. The ML method samples the multidimensional space of Gibbs free energy parameters and user-defined physical constraints into a database of millions of PDs in order to identify the target material properties. The method presented herein is 1000x to 100,000x faster than currently available approaches, and defines a new paradigm on the quantification of properties of materials and devices. As an example application, the developed methodology is combined with the most widely used thermodynamic models – regular solution, Redlich-Kister, and sublattice formalisms– to infer the properties of materials for lithium-ion battery applications, reconstructing without human bias, well-established CALPHAD formulations while identifying previously missed stable and metastable phases and associated properties.

2:20 PM
Thermodynamic Modeling with Uncertainty Quantification and its Implications for Intermetallic Catalysts Design: Application to Pd-Zn-Based Gamma-Brass Phase: Rushi Gong¹; Shun-Li Shang¹; Griffin Canning¹; Robert Rioux¹; Michael Janik¹; Zi-Kui Liu¹; ¹The Pennsylvania State University
    Pd-Zn-based intermetallic catalysts with γ-brass lattice show encouraging combinations of activity and selectivity on well-defined catalytic ensembles. A larger variety of ensembles are accessible if a suitable choice of the third element (M = Au, Ag, Cu, Ni, or Pt) is introduced. In the present work, thermodynamic descriptions of the Pd-Zn system and Pd-Zn-M γ-brass phase have been established using the computational thermodynamics (i.e., the CALPHAD) approach with uncertainty quantification (UQ) through the statistical distribution of model parameters during the Markov Chain Monte Carlo optimization. Activity and selectivity are sensitive to the change of ensembles from Pd monomers to trimers or Pd-M-Pd, which are related to the site occupancies of Pd and M in γ-brass phase. Site occupancies and their UQ, predicted from modeling and compared with the present experiments, are essential to determine ensembles as a function of composition, thus achieving atomic control of catalytic ensembles of intermetallic surfaces.

2:40 PM
Efficient Phase Diagram Determination via Sequential Learning: Theresa Davey¹; Brandon Bocklund²; Zi-Kui Liu³; Ying Chen¹; ¹Tohoku University; ²Lawrence Livermore National Lab; ³Pennsylvania State University
    Phase diagrams are a fundamental tool for materials design, but thorough experimental exploration of the composition and temperature space is challenging, expensive, and time consuming. Despite recent strides in the development of theoretical methods for phase diagram calculations, experimental investigation remains essential to validate the predictions for a high accuracy description. The Gibbs energy of liquid and disordered phases is challenging to obtain directly from first-principles calculations, so other methods are required to fully elucidate the correct phase diagram topology. Considering the quantified uncertainty of the phase diagram, a sequential learning approach is developed to systematically add data in regions of highest uncertainty. Fictitious experimental data is generated and used in automatically optimising a principles-only thermodynamic database. The convergence of the system is examined as various “experimental” data sets are used, demonstrating the selection of an efficient experimental pathway to a high accuracy description.

3:00 PM
A Feature-rich Approach to the Characterization of High Temperature, Sulfate-induced Corrosion of Advanced Alloys: David Poerschke¹; Atharva Chikhalikar¹; ¹University of Minnesota
    Hot corrosion of alloys is caused by a variety of environmental contaminants including sulfates, chlorides, and oxides that deposit on alloy and coating surfaces. The corrosion pathway is strongly influenced by the nature of this deposit and the degree to which the deposit melts and spreads. Corrosive degradation manifests as local and global acceleration of the oxidation rate leading to increased thermally grown oxide (TGO) thickness, roughening of the alloy-TGO interface, and other localized attack. This work has developed automated image analysis tools and a statistical analysis workflow to quantify the effects of changes in alloy chemistry, corrosive deposit composition, and oxidation atmosphere on the prevalence of specific hot corrosion features. Probability distributions were analyzed for key features, and a power spectral density analysis was used to understand the frequency and intensity of interface roughening. This analysis framework enables the generation of large, statistically-meaningful datasets to train alloy design models.

3:20 PM Break

3:40 PM
Bayesian Calibrated Yield Strength Model for High-entropy Alloys: Xin Wang¹; Wei Xiong¹; ¹University of Pittsburgh
    Yield strength prediction is vital in new alloy design, and the solid solution strengthening effect is essential for accurately predicting yield strength. However, the conventional solid solution strengthening model is less accurate for the high-entropy alloys (HEAs) since the HEAs nature of the complex concentrations breaks the concept of solutes and solvents. In this work, we conducted a Bayesian model calibration to optimize the solid solution strengthening model parameters and quantified the model uncertainties based on a relatively large dataset collected from the literature. The accuracy of our model is higher compared with existing models. Moreover, we also evaluated the confidence in each model parameter. The parameter uncertainties were further discussed to identify the knowledge gaps in the physics-based understanding of the strengthening effect in HEAs.

4:00 PM
Using Scalable Multi-Objective Bayesian Optimization to Develop Aluminum Scandium Nitride Molecular Dynamics Force Fields: Jesse Sestito¹; Michaela Kempner²; Tequila Harris²; Eva Zarkadoula³; Yan Wang²; ¹Valparaiso University; ²Georgia Institute of Technology; ³Oak Ridge National Laboratory
    Scandium (Sc) doped aluminum nitride (AlScN) exhibits improved piezoelectric properties. To fine tune the material properties for design purposes, an atomistic level understanding of the structure-property (S-P) relationships is needed. Molecular dynamics can be used to understand the S-P relationships. However, the limited availability of force fields has been a challenge for property predictions. In this work, a force field calibration method using scalable multi-objective Bayesian optimization is presented. Optimizations with three, six, and eight objectives are applied to calibrate AlScN force fields based on piezoelectric characteristics, modulus of elasticity, and lattice parameters at different doped levels. The performances of the different force fields are compared, and the performance of the higher dimensional objective problems is discussed. The highly scalable molecular dynamics force field development method is successfully implemented, resulting in the creation of several aluminum scandium nitride molecular dynamics force fields for piezoelectric applications at varying Sc dope levels.

4:20 PM
Uncertainty Quantification of Constitutive Models in Crystal Plasticity Finite Element Method: Anh Tran¹; Hojun Lim¹; ¹Sandia National Laboratories
    Uncertainty quantification (UQ) plays a major role in the verification and validation of computational engineering models and simulations, cements trust, and establishes the predictive capability of forward models. Crystal plasticity finite element method (CPFEM) has been widely used as one of a few ICME toolboxes that allow numerical predictions from microstructure to materials properties and performances. In this work, we apply a mathematically rigorous stochastic collocation method to quantify the uncertainty of the three most commonly used constitutive models in CPFEM, namely phenomenological models (with and without twinning), and dislocation-density-based constitutive models, for three different types of crystal structures, namely face-centered cubic copper, body-centered cubic tungsten, and hexagonal close packing magnesium. Our work not only quantifies the uncertainty of these constitutive models in the stress-strain curve but also analyzes the global sensitivity of the underlying constitutive parameters with respect to the initial yield behavior.

4:40 PM
Uncertainty Quantification of a High-throughput Local Plasticity Test: Profilometry-based Indentation Plastometry of Al 7075 T6 Alloy: Aaron Tallman¹; Denny John¹; Tanaji Paul¹; Arvind Agarwal¹; ¹Florida International University
    The quantification of spatially variable mechanical response in structural materials remains a challenge. Additive manufacturing methods result in increased spatial property variations—the effect of which on component performance is of key interest. To assist iterative design of additively manufactured prototypes, lower-cost benchtop test methods with high precision and accuracy will be necessary. Profilometry-based indentation plastometry (PIP) promises to improve upon the instrumented indentation test in terms of the measurement uncertainty. PIP uses an isotropic Voce hardening model and inverse numerical methods to identify plasticity parameters. To quantify the uncertainty of the PIP test, ninety-nine PIP tests are performed on prepared portions of an Al 7075 plate sample. The profilometry data and the Voce parameter predictions are examined to distinguish contributions of noise, individual measurement uncertainty, and additional set-wide variations. The quantification of material variability in the presence of measurement uncertainty is discussed.