Frontiers of Materials Award Symposium: Machine Learning and Autonomous Researchers for Materials Discovery and Design : Session I
Program Organizers: Keith Brown, Boston University
Thursday 8:30 AM
February 27, 2020
Room: 4
Location: San Diego Convention Ctr
Session Chair: Brian DeCost, National Institute of Standards and Technology
8:30 AM Invited
Adaptive Machine Learning for Efficient Navigation of Materials Space: Prasanna V. Balachandran1; 1University of Virginia
One of our research interests involve development of efficient data-driven strategies for navigating the vast search space of material possibilities. We asked the following question, is there a relationship between ML model quality, utility functions, and the rate at which optimal materials are discovered? Our on-going empirical work appear to indicate that the rate of discovery is dictated by the nuances of the composition–property landscape. Having poor ML models does not equate to poor research outcomes, provided appropriate input descriptors are included that capture the structure-property relationships. Further, utility functions that evaluate the exploration-exploitation tradeoff do not always produce a “winning” search strategy. Examples will be discussed that highlight the non-trivial nature of adaptive machine learning in materials science domain.
9:10 AM Invited
Unraveling Hierarchical Materials using Autonomous Research Systems: Keith Brown1; 1Boston University
Nature realizes extraordinary material properties through the hierarchical organization of polymers from the molecular to the macroscopic scale. Synthetically recapitulating this level of control has been a long-standing challenge as it requires mastery of each scale and an understanding of how to piece these levels together. Practically, however, there are too many distinct material compositions and processing conditions to test using conventional hypothesis-driven research. Thus, new experimental paradigms are needed. Here, we describe our recent progress using advances in machine learning and automated research systems to study hierarchically structured polymers. In particular, we discuss the degree to which experimental research can be accelerated through the combination of automated experimental systems and machine learning to choose experiments. To explore the merits of such autonomous experimental systems, and discover novel mechanical metamaterials, we present a Bayesian experimental autonomous researcher (BEAR) that combines additive manufacturing, robotics, and mechanical characterization to rapidly construct, test, and, study mechanical structures. Using this platform, we study the elastic and plastic mechanics of polymer structures. Critically, we find that the use of a BEAR enables us to discover high performance structures in 60 times fewer experiments than grid-based experimentation. In addition to rapidly developing an understanding of a family of mechanical structures, these experiments provide important lessons regarding how machine learning and automation can accelerate experimental research and mechanical design. Finally, we describe recent efforts to adopt this autonomous research framework at the nanoscopic scale using scanning probes to create and interrogate libraries of polymers. Ultimately, understanding and leveraging the hierarchical arrangements of materials is a grand challenge. Autonomous research systems that span additive manufacturing, machine learning, and advanced characterization have the potential for transformatively advancing the pace of research to meet this challenge.
9:50 AM Break
10:10 AM Invited
Combining Simulation and Autonomous Experimentation for Mechanical Design: Aldair Gongora1; 1Boston University
Additive manufacturing (AM) has increased the complexity with which structures can be designed and fabricated. Computational tools, empowered by the control afforded by AM, have enabled the discovery and realization of structures with enhanced or tailored mechanical performance. However, this approach is limited to mechanical properties that can be reliably predicted using simulation. For properties that cannot be reliably simulated, such as toughness, autonomous experimental research platforms have emerged to explore the design space for high performing structures by combining automated experimentation and active learning. An open question that remains is how to effectively combine simulation, with varying degrees of accuracy and cost, and autonomous experimentation in order to accelerate learning. In this work, we evaluate a series of methods for combining simulation and autonomous experimentation.
10:50 AM Invited
Closing the Loop in Autonomous Materials Development: Kristofer Reyes1; 1University at Buffalo- the State University of New York
Closed-loop, sequential learning is a key paradigm in autonomous materials development. Within this framework, aspects of the materials system under study are modeled, and such models are used to decide subsequent experiments to be run, results of which are fed-back to update models. Research within this nascent field has focused primarily on the modeling or decision-making aspects of this closed-loop. There are, however, other key components of the loop that deserve attention. In this talk, I will focus on two such components. First, I will describe work in autonomous materials characterization, in which rich characterization data such as microscopy images or three-dimensional reconstructions from atom probe tomography are analyzed without human intervention to encode experimental results for use to update models. Second, I will discuss work on prior knowledge formation and elicitation from experts, which is an important “step 0” within this closed-loop framework.
11:30 AM Invited
Bayesian Methods for Concrete Creep Prediction and Learning Optimized Concrete Microstructure Design: Mija Hubler1; 1University of Colorado Boulder
In past years machine learning has been used to update prediction models for the viscoelastic behavior of concrete. Short-term laboratory tests can only inform certain parameters in science and mechanics-based models of the time-dependent behavior of concrete. Once these models have been empirically calibrated through optimization, they provide a poor prediction. Only be introducing additional data in the form on long-term structural measurements or field testing through Bayesian methods could prediction models provide useful long-term estimates of concrete behavior. More recently, machine learning is being used to automate petrography to assess and diagnose the deterioration state of concrete from image data. The most recent advances in these efforts aim to develop microstructure descriptors of concrete which directly correlate to the strength, stiffness, and toughness of the material. Successfully establishing these descriptors will enable the design of printed concrete microstructures for desired properties.