AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: Material Design II
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Anthony Rollett, Carnegie Mellon University; Francesca Tavazza, National Institute of Standards and Technology; Christopher Woodward, Air Force Research Laboratory
Tuesday 2:30 PM
March 1, 2022
Room: 256A
Location: Anaheim Convention Center
Session Chair: Taylor Sparks, University of Utah; Jason Gibson, University of Florida
2:30 PM
DiSCoVeR Algorithm for Identifying Promising Unlikely Candidates for New Materials: Sterling Baird1; Tran Diep1; Taylor Sparks1; 1University of Utah
Machine learning already enables the discovery of new materials by providing rapid predictions of properties to complement slower calculations and experiments. However, a persistent criticism of machine learning enabled materials discovery is that new materials are very similar, both chemically and structurally, to previously known materials. This begs the question “Can machine learning ever learn new chemistries and families of materials that differ from those present in the training data?” Here, we propose the Descending from Stochastic Clustering Variance Regression (DiSCoVeR) algorithm to systematically discover unintuitive and even unlikely yet promising candidates for new materials. The approach leverages clustering algorithms and introduces a loss function penalty for suggesting candidates close to clusters of known materials. Furthermore, we utilize the Earth Movers Distance approach with a modified Pettifor scale to encode chemical similarity in addition to the traditional composition-based features. We show an ability to extrapolate towards unexpected and unusual candidates.
2:50 PM
NOW ON-DEMAND ONLY – Physics Based Analytical Models for the Design of New Metastable Rare-earth Compounds: Prashant Singh1; T. Del Rose1; Guillermo Vazquez2; Raymundo Arroyave2; Yaroslav Mudryk1; 1Ames Laboratory; 2Texas A&M University
Rare earths find uses in many applications due to their vast span of distinctive physical and chemical properties such as permanent magnets and magnetocaloric materials. The application of machine-learning approaches in rare-earth intermetallic design has been sparse, however, due lack of reliable databases. We utilized `in-house’ rare-earth database to train a SISSO (sure independence screening and sparsifying operator) based machine-learning model and developed physically interpretable analytical models to assess the thermodynamic stability of rare-earth compounds. The analytical models were used for extensive exploration of the alloying effect on thermodynamics stability of Ce based cubic Laves phases with MgCu2 crystal structure. Our predictions are confirmed by experiments and provide quantitative guidance for compositional considerations within machine-learning models for discovering new metastable materials. A detailed electronic-structure analysis of Ce-Fe-Cu systems was also discussed to understand the electronic origin of phase (in)stability.
3:10 PM
Optimizing Thermoelectric Compositions to Achieve Extraordinary Properties: Andrew Falkowski1; Taylor Sparks1; 1University of Utah
Discovering new thermoelectric materials requires a careful balance between competing properties. Data-driven approaches offer new avenues for materials discovery, but suffer from systematization bias; that is, a reliance on preselected candidate lists derived from researcher expectations. Fractional optimization shows promise in overcoming this issue by utilizing model knowledge of chemical systems to find optimal compositions with high fractional resolution. Borrowing concepts from neural style transfer, fractional optimization adjusts the fractional components of a compound to maximize or minimize one or more predicted properties. Here, we demonstrate the capabilities of fractional optimization in finding new thermoelectric materials using a compositionally restricted attention-based network trained on non-stoichiometric compounds with dopants of varied concentration. Promising dopants for a variety of chemical systems were identified that balance thermal and electric transport. These systems were optimized to create compositions with predicted properties that are competitive with and exceed the properties of known thermoelectric materials.
3:30 PM
Deep Neural Network Regressor for Phase Fraction Estimation on the High Entropy Alloy System Al-Co-Cr-Fe-Mn-Nb-Ni: Guillermo Vazquez Tovar1; Sourav Chakravarty1; Rebeca Gurrola1; Raymundo Arroyave1; 1Texas A&M University
High Entropy Alloys (HEAs) are composed of more than one principal element and constitute a major paradigm in metals research. A thorough estimation of the phases that form in HEAs given different elemental input is of paramount importance in designing HEAs. Machine Learning presents a feasible and non-expensive method of predicting HEA phase fractions. A Deep Neural Network (DNN) model is developed for the system Al-Co-Cr-Fe-Mn-Nb-Ni, using a dataset of nearly a million points generated via Thermo-Calc. A list of features was then compiled, heavily influenced by previous works and freely available databases. A feature significance analysis expands the current knowledge on phase-elemental properties relations. The final regressor model shows high performance with coefficient of determination values above 0.98 when identifying the most recurrent phases: BCC, FCC, LAVES, and SIGMA. The DNN is then used as a faster and easier to implement surrogate model for optimization problems.
3:50 PM
A Novel Approach for Rapid Alloy Development Leveraging Machine Learning: Nhon Vo1; Ha Bui2; 1NanoAl LLC; 2Amatrium Inc.
Alloy development, starting with trial-and-error method, has advanced significantly thanks to multi-scale computer simulations and high-throughput experiments. Recently, with vast amounts of data generated from research in the last decades together with advancements in machine-learning, a powerful methodology for alloy development has emerged. In this work, we demonstrate how a particular machine-learning methodology can be embedded in the traditional alloy development workflow to further shorten the timeline, eliminate unnecessary experiments, and reduce cost. We show that several material properties - including some which are difficult or time-consuming to measure or calculate, such as thermal conductivity, fatigue strength or machinability - can be predicted with high accuracy in a matter of seconds. The same methodology can be adapted to multiple families of metallic materials without model rebuilding or retraining. Lastly, we outline some limitations to be addressed in the future to fully capture the power of machine-learning in alloy development.
4:10 PM Break
4:30 PM
Accelerated Genetic Algorithm via a Pre-trained Crystal Graph Convolutional Neural Network: Jason Gibson1; Richard Hennig1; 1University of Florida
Genetic algorithms (GA) are a key tool in crystal structure prediction which enables novel material discovery. However, the speed and scope of the GA is hindered by the computational cost of the density functional theory (DFT) calculations needed to evaluate the formation energy. In this study, we first acquired a dataset of approximately 130k relaxed crystal structures and formation energy calculations. The atoms within these structures were perturbed to create a training set of approximately 2 million crystal structures with the target value of each structure set to the structures relaxed formation energy. This training set was then used to pre-train a crystal graph convolutional neural network (CGCNN) to predict the relaxed formation energy of GA produced structures enabling a filtration process that removes high energy structures prior to DFT evaluation. As the GA progresses the CGCNN’s fidelity is increased by tuning the model on the DFT evaluated structures.
4:50 PM
Balancing Data for Generalizable Machine Learning to Predict Glass-forming Ability of Ternary Alloys: Yi Yao1; Timothy Sullivan1; Feng Yan1; Jiaqi Gong1; Lin Li1; 1University of Alabama Tuscalosa
Machine Learning has thrived on the emergence of data-driven materials science. However, the materials datasets acquired at existing research efforts have significant imbalance issues. This paper investigated the data imbalance for the glass-forming ability of ternary alloy systems, which consists of abundant, low-fidelity high-throughput data, and sparse, high-fidelity traditional experimental data. We demonstrated a new method to handle the data imbalance and trained artificial neural network (ANN) models on the original vs. balanced datasets. The ANN model trained on the balanced dataset solved the overfitting issue suffered by the model trained on the original dataset. More importantly, the generalizability in predicting the new alloy system was improved in the data-balanced model, evidenced by the leave-one-out validation. Our work highlights the importance of handling data imbalance in material datasets to solve the overfitting issues of machine learning models and further enhance generalizability in predicting the characteristics of the new material systems.
5:10 PM
Data Driven Approach to Design/Discover Intercalating Ions and Layered Materials for Metal-ion Batteries: Shayani Parida1; Avanish Mishra1; Arthur Dobley2; C Barry Carter3; Avinash Dongare1; 1University of Connecticut; 2EaglePicher Technologies LLC; 3CINT, Sandia National Laboratories
This study utilizes data-driven machine learning methods to find alternative 2D materials and intercalating ions beyond Li for metal-ion batteries with high-power efficiencies. A dataset is constructed by performing first-principles density functional theory (DFT) simulations to estimate theoretical capacities and voltages by calculating the binding energies of metal ions on 2D materials. A tree-based regression model is developed to predict the binding energies with unprecedented accuracy. The model suggests that atomic electronegativities and number of valence electrons bear a strong correlation with binding energy. The generated dataset also provides insight into structural accommodations that can be expected upon ion intercalation on various 2D materials. Additionally, a binding energy and structural deformation-based classification model is developed to screen anode materials for the next-generation batteries. The model selects intercalating ion and 2D material pairs that are suitable for batteries, based on calculated voltage and volumetric changes in the 2D material upon intercalation.