Algorithm Development in Materials Science and Engineering: ML Algorithms and Their Applications II
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Solidification Committee
Program Organizers: Mohsen Asle Zaeem, Colorado School of Mines; Mikhail Mendelev, KBR; Garritt Tucker, Colorado School of Mines; Ebrahim Asadi, University of Memphis; Bryan Wong, University of California, Riverside; Sam Reeve, Oak Ridge National Laboratory; Enrique Martinez Saez, Clemson University; Adrian Sabau, Oak Ridge National Laboratory

Thursday 8:30 AM
March 3, 2022
Room: 253A
Location: Anaheim Convention Center

Session Chair: Enrique Martinez Saez, Clemson University; Sam Reeve, Oak Ridge National Laboratory

8:30 AM  
Designing Thin Film Microstructures via Neuroevolution Guided Time-dependent Processing Protocols: Saaketh Desai1; Remi Dingreville1; 1Sandia National Laboratories
    Designing next generation thin films, tailor-made for specific applications, relies on the availability of robust processing-structure-property relationships. Traditional structure zone diagrams are limited to low-dimensional mappings, with machine-learning methods only recently attempting to relate multiple processing parameters to the final microstructure. Despite this progress, structure-processing relationships are unknown for processing conditions that vary during thin film deposition, limiting the range of microstructures and properties achievable. In this talk, we discuss how to use neuroevolution, a reinforcement learning algorithm, to design time-dependent processing protocols that achieve tailor-made microstructures. We simulate deposition of a binary-alloy thin film via a phase-field model, where deposition rates and diffusivities are controlled via neuroevolution. Our neuroevolution-guided protocols achieve well-known microstructures with lateral and vertical concentration modulations, as well as more complex, hierarchical microstructures previously not described in simple structure zone diagrams. Our algorithm provides insight to experimentalists looking for additional avenues to design novel thin-film microstructures.

8:50 AM  
Application of Vision Transformers in Tomography Image Segmentations of AM Parts: Saber Nemati1; Les Butler1; Shengmin Guo1; 1Louisiana State University
    Constructing intelligent networks for automatic and fast image segmentation of tomography images has gained a lot of attention during the last few years. Different variants of Convolutional Neural Networks (CNNs) have shown great performance in supervised segmentation using very limited training data. However, the challenge of improving accuracy and speed especially during the training phase is still ongoing for real-time purposes. In this research, Vision Transformers (ViT) are investigated for improving prediction accuracy and computational cost. The effectiveness of ViT is demonstrated using two different sets of tomography images. The results show improvement in the training speed and attention to global features in Vision Transformers, which makes them a possible candidate for in-situ process monitoring in materials engineering applications.

9:10 AM  
Neural Network Models of Phase Field Simulations: Haiying Yang1; Michael Demkowicz1; 1Texas A&M University
    We explore the potential for neural networks (NNs) to predict two aspects of microstructure evolution, as represented in simple phase field simulations. First, we train NNs to predict the evolution of microstructures under pre-specified, spatially and temporally varying temperature and bias fields. This problem is amenable to solution by existing NN training methods because each input maps to a unique output. Next, we train NNs to solve the inverse problem, namely: what externally applied temperature and bias fields are required to guide microstructure evolution to a pre-specified target state? In this problem, inputs do not map to unique outputs, necessitating an innovative approach to NN training. Our work suggests that NNs may be useful for finding optimal processing parameters for achieving a desired microstructure.

9:30 AM  
Random Forest Regressor Models for the Prediction of Novel Alloy Corrosion Performance: Bonita Goh1; Yafei Wang2; Phalgun Nelaturu2; Thien Duong3; Dan Thoma2; Jason Hattrick-Simpers4; Santanu Chaudhuri3; Adrien Couet2; 1University of Wisconsin Madison; 2University of Wisconsin Madison; 3Argonne National Laboratory; 4National Institute of Standards and Technology
    The FeCrMnNi High Entropy Alloy space shows promise to yield compositions that possess the set of desired properties for next-generation molten salt-based energy systems. However, the combinatorial quarternary composition space is quasi-infinite, which poses challenges to alloy optimization. Using a unique and relatively large set of corrosion data obtained from a high-throughput high-temperature corrosion testing platform developed to train machine learning models, we present Random Forest Regressor models for predicting corrosion performance metrics (eg. dissolved corrosion product concentration in the salt) based on an input vector parametrizing the alloys’ physical properties . The strength of this approach is that importance score rankings of alloy physical descriptors can help to affirm observations of novel corrosion mechanisms. In addition, the model facilitates backwards-projection to determine compositions that have yet to be tested but show promise, helping us efficiently search the composition space for promising alloy corrosion resistance.

9:50 AM Break

10:10 AM  
Predicting Temperature-dependent Oxide Redox Reactions with Machine-learning Augmented First-principles Calculations: Josť Garrido Torres1; Vahe Gharakhanyan1; Tobias Hoffmann Eegholm1; Nongnuch Artrith1; Alexander Urban1; 1Columbia University
    The experimental characterization of high-temperature redox chemistry is challenging and requires specialized equipment. CALPHAD simulations can be an alternative but require assessed phase diagrams for the system of interest. First-principles calculations, on the other hand, can provide robust estimates of redox potentials at zero Kelvin without experimental input, but simulating redox reactions at high temperatures is computationally demanding and often too approximate. Here, we discuss initial progress towards the efficient computational prediction of high-temperature redox chemistry using machine-learning augmented first-principles calculations. We show that a combination of results from zero-Kelvin density-functional theory (DFT) calculations and a machine-learning model trained on temperature-dependent reaction free energies allows predicting reduction temperatures of metal oxides. An initial application to crystalline binary and ternary oxides demonstrates that the temperature dependence of the free energy can be cross-learned, removing the limitation to compounds that have previously been thermodynamically assessed.

10:30 AM  
Using Machine Learning Methods to Decode VOx Diffractograms: Saaketh Desai1; Suvo Banik2; Haidan Wen2; Subramanian Sankaranarayanan2; Remi Dingreville1; 1Sandia National Laboratories; 2Argonne National Laboratory
    Understanding the effect of structural defects on the metal-to-insulator transformation in vanadium oxide is critical to tuning the transformation for various applications. Diffractograms offer a non-intrusive way of characterizing defects, but can be challenging to interpret comprehensively, limiting our understanding of structure-property relationships. In this talk, we discuss how to employ machine learning methods to identify critical features of diffractograms that correlate with defect statistics and structure phases in VOx structures. We compute X-ray and electron diffraction patterns for stable and metastable VOx structures, whose defects are captured via one-point and two-point statistics such as defect density and defect pair correlation function. Key features of diffractograms are obtained via non-negative matrix factorization, and a Gaussian process model predicts one-point and two-point statistics using these features. We discuss how our automated workflow comprehensively analyzes diffractograms, accurately capturing defect statistics for simulated and experimental diffraction patterns, reducing the need for additional characterization.