First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): Machine Learning/Deep Learning in Materials and Manufacturing 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
Tuesday 9:30 AM
April 5, 2022
Room: Riverboat
Location: Omni William Penn Hotel
9:30 AM Break
10:00 AM Invited
Enabling Rapid Validation and Dynamic Standardisation of Advanced Manufactured Parts: Gareth Tear; Jose Videira1; James Bird1; 1Synbiosys
We are treating materials of the 21st century in the same way as the predictable materials of the 19th century. For advanced materials such as additively manufactured or fibre composite parts, the material and the structure are one and the same thing. This makes manufacturing unpredictable, unrepeatable, and ultimately more complex. Currently, validation and signing off of components for safety critical parts is prohibitively expensive in both time and cost for most engineering applications. Synbiosys presents their AI validation platform. We discuss how we enable in rapid time: (a) Trust-free verification and validation of advanced material components, (b) Dynamic standardisation, where standards are set by the design requirements and design tolerances are fed down the supply chain, ensuring that each supplier knows what they need to achieve, and how to achieve it, (c) Risk quantification and reduction for decision makers, providing all information in one place.
10:30 AM
Property Optimization of Multifunctional Materials with Complex Parameter Spaces: Kevin Ferguson1; Ayesha Abdullah; Eric Harper2; Levent Kara1; Michael Bockstaller1; Larry Drummy2; 1Carnegie Mellon University; 2Air Force Research Laboratory
Multifunctional material systems present rich parameter spaces that cannot be addressed through traditional iterative material design methods. Experiments and simulations attempting to characterize bulk-material properties (e.g., mechanical or optical performance) over complex parameter spaces are prohibitively expensive. A multi-component material testbed comprised of self-assembled polymer-grafted nanoparticles is introduced. Rather than performing a bulk simulation with a large number of interactions between each pair of multi-component particles, a coarse-grained simulation is proposed. This simplifies the large number of polymer interactions into a single interaction potential whose parameters are optimized via machine learning on empirical data, significantly reducing the complexity of the simulation but maintaining the ability to match experimental results. Such a coarse-grained model can be used to rapidly generate material property data with systematically varied design parameters, augmenting the sparse experimental dataset for the purpose of optimizing target material properties.
10:50 AM
Data Driven Microstructure Evolution: Adjusting Growth Speed and Anisotropy: Joseph Melville1; Amanda Krause1; Joel Harley1; Weishi Yan1; Lin Yang1; Michael Tonks1; 1University of Florida
Microstructural grain growth simulations are used to predict microstructure evolution over time. Conventional grain growth simulations are based on many available techniques, including phase field, cellular automata, and Monte-Carlo-Potts methods. Yet, the rigid physics-based assumptions of these methods have difficulty capturing some real world grain growth behaviors, such as abnormal grain growth. As a result, it can be difficult to adapt these simulations with experimental knowledge. This presentation discusses a physics-informed deep learning framework for microstructural grain growth simulation. This hybrid model enables us to train microstructure evolution based on a combination of physical laws and data (from simulations or experiments). This presentation specifically highlights how this framework learns to adjust growth speed and anisotropy from data. In addition, due to our integration of physics, this adaptive process does not require huge amounts of new data.
11:10 AM
Plastic Deformation Reconstruction Based on Acoustic Emission Measurements: Junjie Yang1; Yejun Gu1; Daniel Magagnosc2; Tamer Zaki1; Jaafar El-Awady1; 1Johns Hopkins University; 2Army Research Laboratory
Interpreting acoustic emission (AE) is a non-destructive evaluation (NDE) technique based on the elastic stress waves caused by irreversible deformations in materials. The core challenge is decoding AE signals and quantitatively determining the source, or deformation mechanisms. We adopt variational and machine-learning approaches to decode the complex interconnections between plastic deformation mechanisms in metals and the AE they emit. Surrogate measurements of AE associated with different dislocation mechanism were first obtained from full 3D elastodynamic-field solution incorporated into 3D Discrete Dislocation Dynamic simulations. Data assimilation procedures use the AE signals at individual sensors only, and reconstruct the plastic deformations. We show that the dislocation activities history in 3D can be accurately reconstructed using a single AE sensor. This approach enables the decoding of hard-to-interpret surface AE measurements and reconstruct plastic slip intrinsically in the material during deformation.
11:30 AM
Analysis of High-Speed Impact Behaviour of Al 2024 Alloy Using Machine Learning Techniques: Navya Gara1; Siri S1; Velmurugan R1; Jayaganthan R1; 1IIT Madras
The large deformation of metallic materials subjected to high-speed impact may lead to catastrophic failure of aerospace structural components fabricated using these materials. The present work is focused to analyze dynamic behavior of Al2024 alloy subjected to very high impact velocities using FEA software (LS DYNA)along with Machine Learning (ML) techniques. The transient impact behaviour of the alloy was estimated using modified Johnson- Cook visco-plastic model for a strain rate range of 100-3000/s. The residual velocities and energy absorption characteristics of Al 2024 subjected to high-speed impact estimated through FEA and analytical routes along with experimental data were utilized to train the ML models such as support vector machine, random forest, and deep neural networks for predicting the crashworthiness of structures. The comparative analysis of ML algorithms was made for its predictive accuracy in estimating the dynamic behaviour of Al 2024 alloy.
11:50 AM
Uncovering Atomic Structure-Property Relationships Driving Segregation Energy Behavior: Jacob Tavenner1; Ankit Gupta1; Garritt Tucker1; 1Colorado School of Mines
A detailed understanding of segregation behavior is critical for determining alloy properties. As nanoscale microstructural processing techniques improve, the limits of macroscale segregation analysis become more apparent. However, simulation of atomic segregation behavior remains computationally expensive. To improve computational capabilities and investigate atomic features important to segregation energy, an improved atomic fingerprinting method is implemented using a new descriptor framework known as Strain Functional Descriptors (SFDs). Coupled with modern machine learning (ML) techniques, this fingerprint provides an accurate determination of atomic segregation energy based only on a-priori information, bypassing computationally complex per-particle energy minimization and/or Monte-Carlo methods. From this atomic fingerprint, our understanding of specific structure-property relationships is also improved by a detailed per-particle analysis of segregation energetics. By investigating per-particle relationships and structural variation present across multiple ML training datasets (including bicrystalline grain boundary (GB), polycrystalline, and bulk amorphous simulations) the advantages and shortcomings of each model are identified.