Materials Design through AI Composition and Process Optimization: Session I
Program Organizers: Noah Paulson, Argonne National Laboratory; Tiberiu Stan, Asml; Brandon Bocklund, Lawrence Livermore National Laboratory; Arun Kumar Mannodi Kanakkithodi, Argonne National Laboratory

Monday 2:00 PM
November 2, 2020
Room: Virtual Meeting Room 41
Location: MS&T Virtual

Session Chair: Noah Paulson, Argonne National Laboratory; Tiberiu Stan, Northwestern University


2:00 PM  Invited
Realistic 3D Microstructure Generation via Generative Adversarial Networks: Elizabeth Holm1; Tim Hsu1; William Epting2; Hokon Kim1; Harry Abernathy2; Gregory Hackett2; Anthony Rollett1; Paul Salvador1; 1Carnegie Mellon University; 2U.S. DOE National Energy Technology Laboratory
    Generating large data sets of realistic, statistically equivalent, 3D microstructures is a prerequisite for computational surveys and optimization approaches. Using large-scale 3D microstructural data, a Generative Adversarial Network (GAN) model has been implemented to learn and generate realistic 3D polyphase solid oxide fuel cell microstructures. Nearly limitless microstructural instances can be generated at relatively minimal cost compared to conventional simulation/statistical approaches. Besides being visually similar to the experimental microstructures, the GAN-generated microstructures are statistically similar with respect to geometric and topological metrics (e.g., particle size, surface area, triple-phase-boundary density). Furthermore, performance simulations applied to the GAN-synthetic microstructures result in realistic electrochemical response. This suggests the GAN model is capable of learning and generating microstructures that captures all salient aspects of the target system. Intriguingly, limitations of the GAN in resolving outlier structures not only provide materials insight, but also reveal opportunities to understand and improve the GAN approach itself.

2:25 PM  Invited
Artificial Intelligence for Material and Process Design: Marius Stan1; Noah Paulson1; Debolina Dasgupta1; Jessica Pan2; Joseph Libera1; 1Argonne National Laboratory; 2Princeton University
    Modeling properties and evolution of complex systems requires a comprehensive evaluation of data and model quality. With the volume, variety and rate of data generation continuously increasing, human analysis becomes extremely difficult, if not impossible. Fortunately, recent advances in artificial intelligence (AI) have significantly improved R&D methodologies by emphasizing the role of the human-machine partnership. We discuss the development of “intelligent software” that includes elements of AI such as machine learning and computer vision, coupled with reduced-order modeling and Bayesian statistics. We illustrate the value of the approach using examples of material design and real-time optimization of manufacturing processes.

2:50 PM  
NEW - Polymer Property Prediction and Design through Multi-task Learning : Christopher Kuenneth1; Lihua Chen1; Huan Tran1; Chiho Kim1; Rampi Ramprasad1; 1Georgia Institute of Technology
    Polymers are an important class of materials that display morphological complexity and diversity spanning a huge property space. Machine learning methods have been recently successfully deployed to explore this unknown polymer property space revealing previously unidentified and novel polymers. The training of machine learning models requires a numerical representation of polymers, commonly termed fingerprints, as inputs which are "mapped" to the polymer properties as outputs. Single-task machine learning models learn the mapping between fingerprints and a single property. Contrarily, multi-task models learn the simultaneous prediction of multiple properties including cross-property correlations. Once trained, multi-task models can not only capture polymer properties but also their correlations which can be extracted and verbalized into polymer design instructions. In this work, we developed a multi-task model for 15 different polymer properties. A comprehensive comparison with single-task models demonstrates superiority of the multi-task model. Moreover, cross-property knowledge is extracted and design instructions are demonstrated.

3:10 PM  Invited
Text and Data Mining for Materials Synthesis: Elsa Olivetti1; 1Massachusetts Institute of Technology
    Predictive materials modeling can provide properties of real and virtual compounds and will be available on demand, thereby enabling rapid iteration time in materials design. However, the allure (and necessity) of accelerated discovery that motivates computational materials design is diminished by the prevalent heuristic approaches to materials synthesis and optimization. This delay in moving from promising materials concept to validation, optimization, and scale–up is a significant burden to commercialization. I will describe our work to extract information from peer reviewed academic literature across a range of inorganic solid state materials synthesis approaches. We have demonstrated not only the potential of the natural language processing approach to assemble materials data from the literature, but we have also shown that one can develop hypotheses for what synthesis conditions drive a particular target material outcome using learning approaches.

3:35 PM  Invited
Deep Materials Informatics: Illustrative Applications of Deep Learning in Materials Science: Ankit Agrawal1; 1Northwestern University
    The growing application of data-driven analytics in materials science has led to the rise and popularity of the relatively new field of materials informatics. Within the arena of data analytics, in recent years deep learning has emerged as a game-changing technique, which has enabled numerous real-world applications such as self-driving cars. In this talk, I would present some of our recent works at the intersection of deep learning and materials informatics, for exploring processing-structure-property-performance (PSPP) linkages in materials. Illustrative examples include learning the chemistry of materials using only elemental composition, learning multiscale homogenization and localization linkages in high-contrast composites, and deep adversarial learning for microstructure design. The increasingly availability of materials databases and big data in general, along with groundbreaking advances in data science approaches offers lot of promise to accelerate the discovery, design, and deployment of next-generation materials.

4:00 PM  
Machine Learning Prediction of Glass Properties Informed by Synthetic Data: Kai Yang1; Mathieu Bauchy1; 1University of California, Los Angeles
    Developing novel glasses with new, improved properties and functionalities is key to address some of the Grand Challenges facing our society. Although machine learning offers a unique opportunity to accelerate the discovery of novel glasses with exotic functionalities, it faces several challenges. In particular, the use of machine learning requires as a prerequisite the existence of data that are (i) available, (ii) complete, (iii) consistent, (iv) accurate, and (v) numerous. For instance, although some glass property databases are available, inconsistencies between data generated by different groups render challenging any meaningful application of machine learning approaches. Here, we present a new machine learning framework that simultaneously leverages experimental and simulation-based (synthetic) data by means of Multi-Fidelity Gaussian Process Regression (GPR). We show that our hybrid model systematically outperforms models relying solely relying on experimental data.

4:20 PM  
Statistics-based Microstructural Digital Image Correlation Method for Estimating Ex-situ Strain from Dissimilar Micrographs: Patxi Fernandez-Zelai1; Quinn Campbell1; Yousub Lee1; Michael Kirka1; Sebastien Dryepondt1; 1Oak Ridge National Laboratory
    Digital image correlation (DIC) methods are ubiquitously used throughout engineering for estimating in-situ strain. Ex-situ DIC estimation of strain from deformed micrographs is not possible as there are no persistent trackable features. Two point spatial statistics enable the quantification of spatial patterns in heterogeneous media. Similar to particle tracking methods, computation of two point statistics rely on the use of convolutions suggesting there is a connection between the two. In this talk we present a novel method for estimating strain directly from dissimilar micrographs using a continuum mechanics approach. The proposed method can be interpreted as a statistics-based microstructural digital image correlation method as it operates on image statistics rather than directly on images. A Bayesian bootstrapping framework is proposed for quantifying prediction uncertainty. This method is broadly applicable in a number of settings: materials processing, dynamically impacted materials, and failure analysis, to name a few.

4:40 PM  
Learning Through Domain Knowledge: A Hierarchical Machine Learning Approach Towards the Prediction of Thermoplastic Polyurethane Properties: Joseph Pugar1; Newell Washburn1; 1Carnegie Mellon University
    Incorporating domain knowledge into machine learning techniques for materials design improves predictive capability on small size datasets. Here we propose a hierarchical machine learning (HML) approach to predict bulk mechanical properties. A small library of 18 unique thermoplastic polyurethanes (TPUs) were synthesized to have varying chemical structures, hard segment weight fractions, and functional group indices. The bottom layer of the model is populated in terms of monomer chemical structure, molecular weights, functional indices, and weight fractions. A middle layer, parameterized in terms of the bottom layer descriptors, captures underlying physical properties by incorporating thermodynamic relationships utilizing Group Interaction Modeling (GIM) and measurable experimental values such as surface contact angle (θC), morphology, and quantum descriptors. The domain heavy middle layer is utilized to predict the Young's Modulus for various TPUs which were compared to a test set of various molecular architectures not seen in the training set.