AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Session V
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Francesca Tavazza, National Institute of Standards and Technology; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Kamal Choudhary, National Institute of Standards and Technology; Saaketh Desai, Sandia National Laboratories; Shreyas Honrao, Aionics; Ashley Spear, University of Utah; Houlong Zhuang, Arizona State University
Wednesday 8:30 AM
March 22, 2023
Room: Cobalt 520
Location: Hilton
Session Chair: Ramsey Issa, University Of Utah; Kamal Choudhary, National Institute of Standards and Technology
8:30 AM
Applications of Machine Learning Techniques for Materials Discovery: Suchismita Goswami1; Ichiro Takeuchi2; 1MEST; 2University of Maryland
Considerable efforts have been made to discover novel materials using machine learning techniques, including feature extraction and visualization for identifying similar potential material with known properties. For the identification of novel materials around a user defined compound, neighborhood maps of materials are usually generated employing the dimensionality reduction algorithms, which map the high dimensional features onto two-dimensions. However, a significant difference after the dimensionality reduction in representing the novel materials around a user defined material as compared to the high dimensional featurized space has been observed. Here we implement a different approach that will reduce the observed difference in the featurized and the reduced dimensional space. We employ Matminer and DScribe libraries to featurize crystallographic information files of the ICSD database into numerical feature vectors with JarvisCFID and Sine-Matrix methods, respectively. In this presentation, we will discuss results on the neighborhood maps on transition-metal based ferromagnetic compounds.
8:50 AM
High-dimensional Formulation-based Bayesian Optimization of Dental Composite Resins: Ramsey Issa1; Taylor Sparks1; 1University of Utah
The need for adaptive design via machine learning in dental composite resins (DCRs) has never been greater. With close to 100 tunable parameters made up of resins, fillers, pigments, and initiators the task becomes daunting when only approaching this high dimensional problem using domain knowledge. Thus, utilizing a more efficient design space search is put forward employing a Bayesian optimization technique, optimizing for the compressive strength of DCRs. We constrain the design space to optimize for 16 resins holding fillers, pigments, and initiators constant. At each Bayesian optimization iteration, we generated 50,000 random compositions that were evaluated using a surrogate function to quantify the mean and standard deviation of the predicted compressive strength. An acquisition function was then utilized to maximize the expected improvement of these samples. The samples were synthesized, characterized, and the data was fed back into the model to complete the adaptive design loop.
9:10 AM
Accelerated Discovery of Ultra-high Temperature High Entropy Ceramics by Machine Learning and High Throughput Experiments: Kun Wang1; Yonggang Yan1; 1Alfred University
Machine learning (ML) methods have been successfully applied to predict phase formation and properties of novel materials, such as high entropy materials. However, the application of ML approach in ultra-high temperature high entropy ceramics which are potential high-temperature structural materials in extreme environments such as nuclear reactors and hypersonic environments etc., remains limited researched due to the serious data scarcity as well as data quality issues. Herein, the high-throughput experiments were employed to generate high-quality dataset for ML training, because the experiment was conducted under the same conditions and by the same researcher. The experimental validation is performed to examine the performance of the ML model. In addition, the ML is also applied to discover the most relevant input features with respect to the output properties, giving rise to an inverse understanding of the underlying physical mechanisms. In particular, we discovered an empirical phase classification rule for high entropy diborides.
9:30 AM
A Generative AI Framework for Designing Nanoporous Silicon Nitride Membranes (NPM) with Optimized Mechanical Properties: Ali Shargh1; Gregory R. Madejski1; James McGrath1; Niaz Abdolrahim1; 1University of Rochester
NPMs are ultrathin freestanding materials containing pores with tunable sizes. It is well-accepted that strength of NPM should improve significantly for applying them in life-saving applications like hemodialysis devices. However, traditional optimization methods cannot be used for empirical optimization of NPM strength as the number of tunable manufacturing parameters are too high. We develop an AI framework for design of NPMs with tunable performance. Our framework is trained on a large dataset containing thousands of NPMs with different microstructures that are labeled with strength from finite element simulations. We show that the framework could generate meaningful pore patterns corresponding to optimized strength values that are either absent or higher than initial dataset. Further evaluation of finite element simulations reveal that strength of the AI design is also higher than common commercialized pore pattern thanks to the decrease of stress concentration. We also validate these findings using experiments and atomistic simulations.
9:50 AM
Designing High-Temperature Multicomponent NiTiHfPd SMAs Using Machine Learning: Hatim Raji1; Soheil Saedi1; 1Florida Institute of Technology
Machine learning (ML) has become an attractive approach for designing multicomponent alloys where the design space is extensive. Quaternary NiTiHfPd shape memory alloys (SMAs) are ultra-strong with a high potential for high-temperature actuation and damping application, yet they are significantly less studied compared to the other members of the SMA family. This study aims to accelerate the slow pace of the research on NiTiHfPd SMAs that is mostly arisen from the high cost of the Pd element using a data-mining approach. To this end, a database created by compiling all published and unpublished data and extended through principles of high entropy alloy design is used. An ML algorithm is developed to predict the transformation temperatures (TTs) and hysteresis of NiTiHfPd of certain compositions and assists to understand the trend in the relationship between the compositions and TTs. Several classifiers are employed, and the predicted data is validated by experiment.