First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): AI-Assisted Development of New Materials/Alloys III
Program Organizers: Taylor Sparks, University of Utah; Michael Dawson-Haggerty, Kerfed, Inc.; Elizabeth Holm, University of Michigan; Jin Kocsis, Purdue University; Adam Kopper, Mercury Marine; Benji Maruyama; James Warren, National Institute of Standards and Technology
Wednesday 9:00 AM
April 6, 2022
Room: William Penn Ballroom
Location: Omni William Penn Hotel
Session Chair: Madison Wenzlick, National Energy Technology Lab
9:00 AM Invited
Gaussian Process as a Flexible Machine Learning Toolbox with Uncertainty Quantification for Solving Inverse Problems in Process-Structure-Property Relationship: Anh Tran1; Tim Wildey1; 1Sandia National Laboratories
Process-structure property is the celebrated hallmark of materials science. Numerous ICME models have been developed to enable accurate numerical predictions, where uncertainty plays a key role in verification and validation. While forward models are widely established, inverse problems are less so due to the expensive computational cost of the forward models. The Gaussian process is a powerful and flexible machine learning toolbox that offers a broad possibility for machine learning in materials science applications, including data-efficient active machine learning approaches such as Bayesian optimization. In this work, we utilize Gaussian process and Bayesian optimization to approximate ICME models and solve deterministic/stochastic inverse problems in process-structure and structure-property relationships. We demonstrate the usefulness of the Gaussian process with two ICME examples: kinetic Monte Carlo and crystal plasticity finite element models.
9:30 AM
Machine-Learning Force-Field to Develop and Optimize Multi-Component Alloys: Anup Pandey1; 1Los Alamos National Laboratory
The empirical molecular dynamics (MD)-based atomistic study of multi-component alloys (MCAs) is limited to fewer candidates due to the lack of accurate force fields. It is almost impossible to explore the vast configurational space of MCAs using computationally demanding ab initio methods such as DFT. Machine-learning learning force-fields (ML-FF) have opened up an avenue for the automated search of novel MCAs with desired properties by combining theory and experiments. ML-FF can be utilized for a high-throughput exploration of the alloy components, thereby being a suitable alternative for conventional force fields that have transferability issues. We will discuss the application of ML-FF in optimizing and developing novel refractory high entropy alloys (RHEAs) with desired mechanical, elastic, and thermal properties.
9:50 AM
Designing Thin Film Microstructures Using Genetic Algorithm Guided Time-Varying Processing Protocols: Saaketh Desai1; Remi Dingreville1; 1Center for Integrated Nanotechnologies, Sandia National Laboratories, Albuquerque, NM, USA
Designing next generation thin films, tailor-made for specific applications, relies on the availability of robust processing-structure-property relationships. Traditional structure zone diagrams (SZDs) are limited to low-dimensional mappings, with machine-learning methods only recently relating 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/properties achievable. In this talk, we discuss how to use genetic algorithms (GAs) to design time-dependent processing protocols that achieve tailor-made microstructures. We simulate physical vapor deposition of a binary-alloy thin film via a phase-field model, where deposition rates and diffusivities of the deposited species are controlled via the GA. Our GA-guided protocols achieve targeted microstructures with lateral or vertical concentration modulations, as well as more complex, hierarchical microstructures previously not described in simple SZDs. Our algorithm provides insight to experimentalists looking for additional avenues to design novel thin-film microstructures.
10:10 AM Break
10:40 AM
Gaussian Process Regression Modelling of Superalloy Microstructure: Patrick Taylor1; Gareth Conduit1; 1University of Cambridge
Gaussian process regression is used to model the phase compositions of nickel-base superalloys. The method makes use of a physically informed kernel and a novel Bayesian approach to ensuring predictions are consistent when summed across phases. The GPR model delivers good predictions for laboratory and commercial superalloys with R2>0.8 for all but three components and R2=0.924 for the γ' phase fraction. It also outperforms CALPHAD for predictions on four benchmark SX-series superalloys. Furthermore, unlike CALPHAD the GPR model quantifies uncertainty in predictions, can be retrained as new data becomes available, and can be used to identify outliers in the initial training dataset.
11:00 AM
Experimental Validation of Materials Discovered by Autonomous Intelligent Agents: Carolyn Grimley; Joseph Montoya1; Muratahan Aykol1; Jens Hummelshøj1; Richard Padbury2; 1Toyota Research Institute; 2Lucideon
One of the biggest challenges facing industry is the identification of new materials to meet emerging industrial demands. However, gaps in advanced materials properties remain due to the intractable challenge of screening a vast material composition search space. To overcome this challenge, Toyota Research Institute (TRI) has developed an end-to-end platform that leverages artificial intelligence-based agents to explore high-dimensional search spaces for new materials. The computational autonomy for materials discovery (CAMD) system uses a combination of past knowledge, surrogate models and heuristic rules to promote cost-effective exploration-exploitation strategies that guide which experiments to perform next, aiding the pursuit of new stable compounds. In collaboration with Lucideon we present a complete workflow, from discovery campaign to materials synthesis, to demonstrate how new inorganic materials can be identified by CAMD, down selected for further analysis, experimentally validated using appropriate in-situ characterization techniques and subsequently, processed via solid state synthesis routes.
11:20 AM
Physics-Constrained, Inverse Design of High-Temperature, High-Strength, Creep-Resistant Printable Al Alloys Using Machine Learning Methods: S. Mohadeseh Taheri-Mousavi1; Florian Hengsbach2; Mirko Schaper2; Greg B. Olson1; A. John Hart1; 1MIT; 2Paderborn University
Aluminum alloys that retain strength at high temperatures have various industrial applications including fan blades of jet engines and pistons of combustion engines. However, additive manufacturing (AM) of these alloys is challenging due to the presence of hot cracking. We demonstrate a physics-constrained, inverse design framework with data generated from integrated computational materials engineering (ICME) techniques to explore the compositional space of Al-Zr-Er-Y-Yb-Ni, and identify an optimal alloy composition achieving maximum predicted strength at a temperature of 250ºC. Using only 40 sampling data with our most efficient machine learning algorithm (a neural network), we predicted a composition with coarsening resistance 3.5X greater than a state-of-the-art printable Al-alloy and 1.75X greater than the maximum outcome among 500,000 random compositions. Mechanical testing of coupons 3D-printed from the optimal composition validated our predictions. The numerical framework provides an efficient and robust pathway for alloy design in tandem with the capabilities of AM.