AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis: On-Demand Oral Presentations
Sponsored by: TMS: Computational Materials Science and Engineering Committee
Program Organizers: Mathew Cherukara, Argonne National Laboratory; Badri Narayanan, University of Louisville; Subramanian Sankaranarayanan, University of Illinois (Chicago)
Friday 8:00 AM
October 22, 2021
Room: On-Demand Room 2
Location: MS&T On Demand
Invited
Machine Learning for Automated Experiment in Scanning Probe and Electron Microscopy: Sergei Kalinin1; 1Oak Ridge National Laboratory
Machine learning and artificial intelligence (ML/AI) are rapidly becoming an indispensable part of physics research. the recent successes in applying ML/AI methods for autonomous systems from robotics through self-driving cars to organic and inorganic synthesis are generating enthusiasm for the potential of these techniques to enable automated and autonomous experiment (AE) in imaging. In this presentation, I will discuss challenges and opportunities of AE in SPM and STEM, ranging from feature discovery to controlled intervention. The special emphasis is made on the rotationally invariant variational autoencoders that allow to disentangle rotational degrees of freedom from other latent variables in imaging and spectral data. Extension of encoder approach towards establishing structure-property relationships will be illustrated on the example of ferroelectric domain walls and plasmonic structures. I further discuss the strategies based on Gaussian Processes for automated experiment, and demonstrate some initial results for AE in SPM and STEM.
Invited
Deep Learning and Uncertainty Quantification for Automated Experiments: Bobby Sumpter1; Ayana Ghosh1; Maxim Ziatdinov1; Sergei Kalinin1; Ondrej Dyck1; 1Oak Ridge National Laboratory
In experimental imaging, rapid feature extraction is critical for conversion of the data streams to spatial or spatiotemporal arrays of features of interest. Deep learning while a powerful approach for feature extraction is often limited by the out-of-distribution drift between experiments, where the network trained for one set of conditions becomes sub-optimal for different ones. This limitation is particularly stringent in the quest to have an automated imaging experiment since retraining or transfer learning becomes impractical. To address this gap, we have recently explored the reproducibility of the deep learning for feature extraction in atom-resolved electron microscopy and demonstrated workflows based on ensemble learning and iterative training that greatly improve feature detection. This approach also enables incorporating uncertainty quantification into the deep learning analysis and rapid automated experimental workflows. In this talk I will present a summary of our recent work.
Prediction of Dynamic Properties of LiF and FLiBe Molten Salts with DeepPot Network Potentials: Alejandro Rodriguez1; Hu Ming1; 1University of South Carolina
Molten salts have attracted interest as a potential heat carrier and/or fuel dissolver in developments of new Gen IV reactor designs. Those containing lithium and fluoride-based compounds are of particular interest due to their affinity to lower melting points of mixtures and their compatibility with alloys. A molecular dynamics study is performed on two popular molten salts, namely LiF (50% Li) and FliBe (66% LiF - 33% BeF2), to predict properties, namely density, specific heat, thermal conductivity, and shear viscosity. Due to the large possibilities of atomic environments, we employ training using the Deep Potential (DeepPot) neural networks to learn from large DFT datasets of 118,115 structures with 70 atoms each for LiF and 222,903 structures with 91 atoms each for FLiBe molten salts. These networks are then deployed in fast molecular dynamics to predict dynamic properties that become apparent at longer time-scales, and are otherwise difficult to achieve with man-made potentials, ab-initio, or with experiments.
Understanding the Composition–property Relationship of Glasses Using Interpretable Machine Learning: Ravinder Bhattoo1; Suresh Bishnoi1; Mohd Zaki1; N. M. Anoop Krishnan1; 1Indian Institute of Technology Delhi
The property of inorganic glass is significantly affected by its stoichiometry. Therefore, understanding the composition–property relationship is key for developing novel inorganic glasses. Herein, we use a glass database (>450,000 glass compositions) with up to 232 glass components to train XGBoost (Extreme Gradient Boosting) models for 25 glass properties (including optical, physical, electrical, and mechanical properties). Further, we use SHAP (Shapely additive explanations) to determine each input glass component’s role in controlling the glass property quantitatively. The SHAP analysis reveals a strong interdependence among the glass components for properties like liquidus temperature and glass transition, whereas no such interdependence for properties like density. While some of this interdependence can be explained as “boron anomaly” and “mixed modifier effect”, the others need further exploration. Thus, our work is critical in understanding the component–structure–property relationship of inorganic glasses and discovering novel inorganic glasses.