Materials Design through AI Composition and Process Optimization: Session II
Program Organizers: Noah Paulson, Argonne National Laboratory; Tiberiu Stan, Asml; Brandon Bocklund, Lawrence Livermore National Laboratory; Arun Kumar Mannodi Kanakkithodi, Argonne National Laboratory
Tuesday 2:00 PM
November 3, 2020
Room: Virtual Meeting Room 41
Location: MS&T Virtual
Session Chair: Arun Kumar Mannodi Kanakkithodi, Argonne National Laboratory; Brandon Bocklund, Pennsylvania State University
2:00 PM Cancelled
Autonomous Materials Discovery at the Beamline: A Gilad Kusne1; 1NIST
The last few decades have seen significant advancements in materials research tools, allowing researchers to rapidly synthesis and characterize large numbers of samples - a major step toward high-throughput materials discovery. Machine learning has been tasked to aid in converting the collected materials property data into actionable knowledge, and more recently it has been used to assist in experiment design. In this talk we present the next step in machine learning for materials research - autonomous materials research systems. We present a fully autonomous, closed-loop measurement system in control of X-ray diffraction measurement equipment at the beamline. The system reduces the number of measurements required for phase mapping to 10 % and exploits phase mapping results for material optimization, resulting in the discovery of a novel nanocomposite phase change material. The machine learning algorithm capitalizes on prior knowledge in the form of physics theory and external databases to accelerate discovery.
2:25 PM Invited
High-fidelity Accelerated Design of High-performance Electrochemical Systems: Rachel Kurchin1; Alan Edelman2; Viral Shah3; Chris Rackauckas3; Bryce Meredig4; Venkat Viswanathan1; 1Carnegie Mellon University; 2Massachusetts Institute of Technology; 3Julia Computing; 4Citrine Informatics
Two efforts to curb greenhouse gas emissions involve seeking an electrochemical alternative to the Haber-Bosch process and increasing energy density of lithium-ion batteries to enable electric aviation and trucking. Novel materials can address both these challenges; however, evaluations of candidate material systems in-silico and experimentally are limited to small design spaces and low-fidelity screening that fail to model realistic operating conditions.I will present a collaborative effort spanning academia and industry that aims to alter this paradigm by enabling rapid, high-fidelity screening of large numbers of electrochemical functional materials for use in new energy technologies. This approach will be demonstrated for a model problem in electrocatalysis and shown to extend for battery electrolyte design. We are advancing state-of-the-art along several fronts, including neural differential equations and graph convolutional neural networks for fast evaluation, and efficient global optimization over the chemical space. Experimental validation will be done for high-value candidates.
2:50 PM Invited
Investigating Crystallographic Texture Control Using Laser Powder-bed Fusion Additive Manufacturing: Joseph Pauza1; Anthony Rollett1; 1Carnegie Mellon University
Additive manufacturing of parts fabricated from structural materials occurs on a much finer scale than many other popular manufacturing techniques. The ability to make processing decisions at this scale offers new and promising paths for microstructure control and design. A major component of microstructure in structural materials is crystallographic texture. Parts produced by laser powder bed fusion additive manufacturing have been observed to develop a range of textures during fabrication. The strength of these textures is variable and dependent on a variety of processing factors. The driving physics of these textures is well understood and can be used to inform the processing decisions made during fabrication. We present a study of the modification of laser powder-bed fusion processing parameter to tune crystallographic texture within Inconel 718 parts. Experimental testing is undertaken to understand the mechanical response of the texture control and its impact on overall part performance.
3:15 PM Invited
Multi-information Source Batch Bayesian Optimization of Alloys: Raymundo Arroyave1; 1Texas A&M University
ICME methods and combinatorial materials synthesis/characterization constitute the dominant paradigms for materials development. Unfortunately, they suffer from significant limitations: ICME methods tend to be sequential in nature and are limited by the computational costs of models used to build process-structure-property (PSP) relationships. Combinatorial methods, on the other hand, are "open loop" and are incapable of providing recommendations on the next action to take once information has been acquired. Here, we present a new framework that aims to incorporate the advantages of both paradigms while addressing all their weaknesses. We demonstrate a Multi-Information Source Batch Bayesian Optimization (BO) framework capable of integrating multiple models and information sources at once in order to optimally explore and exploit a materials design space. More importantly, our approach is capable of carrying out this Bayesian-optimal exploration/exploitation in batch mode. This overcomes the major limitation of sequential BO, enabling considerable order-of-magnitude speedups in materials design.
3:40 PM
Accelerating the Discovery of New DP-steel Using Machine Learning-based Multiscale Materials Simulations: Abdallah Chehade1; Tarek Belgasam2; Georges Ayoub1; 1University of Michigan; 2University of Benghazi
While it is of high interest for the transportation industry to design and discover different grades of DP steels exhibiting desirable mechanical properties, this requires exploring a large number of DP steel microstructure combinations. Expensive trial-and-error-based experimentations and multiscale materials simulations are two conventional approaches that have been widely adopted in the field of materials design and discovery. A Gaussian process is developed to accelerate the discovery of the mechanical properties of different DP steels by evolving the microstructure parameters using a limited number of numerical simulations (using a multiscale materials model). The proposed Gaussian process not only accelerates the prediction of the desired mechanical properties of millions of multiscale materials simulations but also offers uncertainty quantification around its predictions. The proposed framework combining multiscale simulations and the Gaussian process is used to discover the microstructural design of DP steel with maximum tensile toughness.
4:00 PM
MeltNet: Predicting alloy melting temperature by machine learning: Pinwen Guan1; Venkat Viswanathan1; 1Carnegie Mellon University
Thermodynamics is fundamental for understanding and synthesizing multi-component materials, while efficient and accurate prediction of it still remain urgent and challenging. As a demonstration of the "Divide and conquer" strategy decomposing a phase diagram into different learnable features, quantitative prediction of melting temperature of binary alloys is made by constructing the machine learning (ML) model "MeltNet" in the present work. The influences of model hyperparameters on the prediction accuracy is systematically studied, and the optimal hyperparameters are obtained by Bayesian optimization. A comprehensive error analysis is made on various aspects including training duration, chemistry and input features. It is found that except a few discrepancies mainly caused by less satisfactory treatment of metalloid/semimetal elements and large melting point difference with poor liquid mixing ability between constituent elements, MeltNet achieves overall success in prediction, especially capturing subtle composition-dependent features in the unseen chemical systems for the first time. The reliability, robustness and accuracy of MeltNet is further largely boosted by introducing the ensemble method with uncertainty quantification. Based on the state-of-the-art underlying techniques, MeltNet achieves a prediction mean average error (MAE) as low as about 120 K, at a minimal computational cost. We believe the present work has a general value for significant acceleration of predicting thermodynamics of complicated multi-component systems.
4:20 PM Cancelled
Using Machine Learning to Classify In Situ Monitored Melt Pool Surfaces in a Powder Bed Fusion Process: Wei Xing1; Yu Zou1; 1University of Toronto
Our work examines the possibility to classify in-situ melt pool images under different laser power levels by the Convolutional Neural Network (CNN) models. Melt pool surface images generated by five levels of peak laser power were employed to train and test eight types of CNN models. The various CNN models were evaluated in terms of classification accuracy and efficiency (accuracy divided by calculation time). Results show that Googlenet has highest accuracy while Alexnet exhibits the highest efficiency. Moreover, visualization methods were employed to analyze the support for CNN classification. Results from Alexnet show that the CNN model extracts features by dividing surface images into melt pool, melt track and powder bed area, and the melt pool area is feature concentrated. Based on above results, two suggestions have been proposed for the melt pool variation detection using CNN models.