Materials Genome, CALPHAD, and a Career over the Span of 20, 50, and 60 Years: An FMD/SMD Symposium in Honor of Zi-Kui Liu: Machine Learning
Sponsored by: TMS Functional Materials Division, TMS Structural Materials Division, TMS Materials Processing and Manufacturing Division, TMS: Alloy Phases Committee, TMS: Integrated Computational Materials Engineering Committee
Program Organizers: Yu Zhong, Worcester Polytechnic Institute; Richard Otis, Jet Propulsion Laboratory; Bi-Cheng Zhou, University of Virginia; Chelsey Hargather, New Mexico Institute of Mining and Technology; James Saal, Citrine Informatics; Carelyn Campbell, National Institute of Standards and Technology
Wednesday 2:00 PM
March 22, 2023
Room: Sapphire L
Location: Hilton
Session Chair: Chelsey Hargather, New Mexico Tech
2:00 PM Invited
The Modern-day Blacksmith: Gareth Conduit1; 1Cavendish Laboratory
We present a machine learning tool, Alchemite, that merges all possible sources of information about a material: simulations, physical laws, and experimental data. Starting from a database of CALPHAD predictions we train a machine learning model that can understand phase behavior, and use this model as an input to predict other material properties. We illustrate the approach with a case study that starts from a training set comprising just ten core results for alloy direct laser deposition, and use Alchemite to augment these with phase behavior & complementary material properties. We ask the model to design the composition and processing parameters of a nickel-base alloy for direct laser deposition, whose properties are then experimentally verified.[1] [1] Probabilistic neural network identification of an alloy for direct laser deposition B.D. Conduit, T. Illston, S. Baker, D. Vadegadde Duggappa, S. Harding, H.J. Stone & G.J. ConduitMaterials & Design 168, 107644 (2019)
2:30 PM Invited
Data-Driven Discovery and Design of Thermoelectric Materials: Christopher Wolverton1; 1Northwestern University
Discovery and design of novel thermoelectrics is particularly challenging, due to the complex set of materials properties that must be simultaneously optimized. Here we discuss our efforts at developing and applying data-driven computational techniques that enable an accelerated discovery of novel thermoelectrics. These techniques involve a combination of high-throughput density functional theory (DFT) calculations, inverse design approaches, and machine learning and artificial intelligence based methods. We discuss several recent examples of these methods: (i) inverse design strategies based on a materials database screening to design a solid with a desired band structure, (ii) inverse design strategies to identify compounds with ultralow thermal conductivity (iii) an effective strategy of weakening interatomic interactions and therefore suppressing lattice thermal conductivity based on chemical bonding principles, and (iv) the development of crystal graph based neural network techniques to accelerate high-throughput computational screening for materials with ultralow thermal conductivity.
3:00 PM Invited
Computational Design of Novel High-Entropy Alloys: Multi-Strengthening Mechanisms vs Neural Network Model: Jaemin Wang1; Hyeon-Seok Do1; Byeong-Joo Lee1; 1Postech
Efforts to design novel HEAs have been made using two different approaches, one based on multi-strengthening mechanisms and the other using neural network modeling. In the first approach, the strength of Co-Cr-Fe-Mn-Ni based HEA was improved via various strengthening mechanisms, such as solid solution strengthening, TRIP and precipitation hardening. We used CALPHAD-type thermodynamic calculations: calculation of phase equilibria, free energy difference between FCC and metastable solid solution phases, and precipitation kinetics. In the second approach, first we suggested a novel and robust neural network model to relieve the burden of searching vast compositional space. Then, we inverse-predicted the process condition to obtain HEAs with good mechanical properties. Finally, we conducted experimental verification on the designed HEAs to prove the validity of the model and alloy design method. The strengthening mechanism of the designed HEAs is discussed by analyzing microstructure and calculating the lattice distortion effect using a molecular dynamics simulation.
3:30 PM Break
3:50 PM Invited
Coupling Physics in Data-driven High-temperature Alloys Design via High-throughput CALPHAD: Dongwon Shin1; Jian Peng1; Yukinori Yamamoto1; Michael Brady1; J. Allen Haynes1; Sunyong Kwon1; 1Oak Ridge National Laboratory
It was recently demonstrated that augmenting raw experimental high-temperature alloy datasets consisting of only elemental compositions and measured properties with CALPHAD-derived synthetic features can efficiently introduce physics into data science. We will present a modern data-driven alloy design workflow that streamlines experimental data curation, high-throughput CALPHAD data augmentation, quantitative correlation analysis for feature selection, surrogate model training, predicting properties of hypothetical alloys, and rapidly screening alloy compositions worthy of experimental validation. We will also discuss challenges and opportunities to further improve the proposed alloy design workflow. This research was supported by the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U.S. Department of Energy (DOE) and the Vehicle Technologies Office Materials Program, U.S. DOE.
4:20 PM Invited
Data-driven Modelling of Metallurgical Processes – A Case Study on BOF Process: Hongbiao Dong1; 1University of Leicester
Steel manufacturing is a long and complicated process including stages of BF ironmaking, BOF steelmaking, refining, casting and rolling; thousands of processing parameters can potentially influence mechanical properties of final products. Recently, significant progress has been made in steel industry to develop online monitoring systems, collecting data for process control. Challenges remain in the area of data storage, cross-process data links, erroneous datasets, the correlation between chemistry, process variables and mechanical properties. The development of a data-driven approach through artificial intelligence (AI) techniques enables machine learning techniques to big datasets aiming to provide process–property optimization and identify challenges and gaps in the data. Recently, computational capabilities and algorithmic developments have significantly grown in power and complexity, accelerating process optimization. Addressing large-scale industrial data process–property optimization strategies involve numerous influencing possessing factors but limited data. Herein, an integrated data-driven steelmaking case study is built with the aim of predicting and optimizing the final product composition and quality. Machine learning is used collaboratively with fundamental knowledge of metallurgy process, and feedback into a backpropagation neural network (NN) model. Integrating data mining and machine learning generates reasonable predictions and addresses process efficiencies within the BOF steelmaking process.
4:50 PM Invited
Efficient Exploration of Compositionally Complex Alloys: Raymundo Arroyave1; Brent Vela1; Danial Khatamsaz1; Douglas Allaire1; 1Texas A&M University
In this talk, I will discuss several strategies that we have beendeveloped in my group and with collaborators to accelerate the rate at which the vast Compositionally Complex Alloy (CCA) space can be explored. Specifically, we introduce some examples in which we use novel ML-assisted algorithms to identify the feasible regions in a number of important FCC and BCC CCA systems. We also discuss the incorporation of closed-loop high-throughput frameworks for the exploration of these alloy spaces via computational and experimental means. Some conclusions on the likely high-value regions of these systems will be presented, which hopefully will motivate further work to explore this very complex chemical space.