Frontiers of Materials Award Symposium: Machine Learning and Autonomous Researchers for Materials Discovery and Design : Session II
Program Organizers: Keith Brown, Boston University
Thursday 2:00 PM
February 27, 2020
Room: 4
Location: San Diego Convention Ctr
Session Chair: Keith Brown, Boston University
2:00 PM Invited
Autonomous Research Systems for Materials Development: Benji Maruyama1; Rahul Rao2; Ahmad Islam2; Jennifer Carpena2; Michael Susner2; Kristofer Reyes3; Jay Myung4; Mark Pitt4; 1US Air Force; 2UES Inc.; 3University at Buffalo, The State University of New York; 4The Ohio State University
Autonomous Research Systems like ARES™ are disrupting the research process by using AI and Machine Learning to drive closed-loop iterative research. ARES™ is our autonomous research robot capable of designing, executing and evaluating its own experiments at a rate of up to 100 iterations per day. Previously ARES taught itself to grow carbon nanotubes at controlled rates (NPJ Comp Mat 2016). Here we discuss recent research campaigns on maximizing carbon nanotube growth rates using a Bayesian optimization planner. We also use HOLMES and knowledge gradient descent to introduce advanced decision policies with local parametric models to control nanotube diameter. Implications for nanotube materials development will be discussed. Finally, we have developed a new research robot for additive manufacturing, AM ARES™, which is at the early stages of teaching itself to print structures with unknown inks. We plan to make the AM ARES™ Robot available online so that a broad community of researchers can test concepts and approaches for AI/ML and experimental design as applied to 3D printing, thus building to the larger goal of enhancing citizen science.
2:40 PM Invited
Application of Machine Learning and Federated Big Data Storage & Analytics for Accelerated Additive Process and Parameter Development: Vipul Gupta1; 1GE Research
In laser powder-bed fusion additive manufacturing (LPBF-AM), part design, materials, machine and post-processing parameters are intertwined, and therefore, require iterative multi-level optimization to meet desired part performance. Ongoing work at GE Research is aimed at robust process optimization, thorough qualification and rapid insertion of additive materials. We developed a physics-informed data-driven framework for LPBF-AM that utilizes probabilistic machine learning, intelligent sampling and optimization protocols, coupled with materials science to dramatically accelerate the process development, and also provide multiple optimal solutions to meet a variety of target material properties. Additionally, to address challenges of maintaining process pedigree, storing experimental datasets, and creating user-friendly analytics, we developed a Federated Big Data Storage and Analytics platform, with the ability to link diverse, multimodal data together to enable complex analytics. In this talk, I will discuss these tools and their applications to parameter optimization for alloy screening, build-productivity, non-coventional particle size distribution and layer-thicknesses.
3:10 PM Invited
Design of Halide Perovskites via Physics-informed Machine-learning: Shijing Sun1; 1Massachusetts Institute of Technology, Photovoltaics Research Laboratory
Improving the environmental stability of halide perovskites is a critical challenge in perovskite solar cell development.1 Despite the remarkable photovoltaic performances,2 methylammonium lead iodide (MAPbI3) is notorious for its heat and moisture instability.3 Intensive research have been put into composition engineering in the past several years, where cation substitutions, e.g. incorporating alkaline metal Cs and small organic ion e.g. formamidinium (FA) into the MAPbI3 lattice, are shown to be among the most effective stabilization strategies.4 However, identifying and optimizing mixed-ion perovskites for reliability in real-world climates is a very challenging task due to the vast composition possibilities and the lack of physics-informed guidance. In this talk I will discuss our recent progress incorporating DFT into a Bayesian optimization algorithm5 to direct the search for novel semiconductors. To effectively design solar materials that are stable under the industrial standard of 85 RH% and 85°C reliability test, we combined the strengths of theory-guided and data-guided methodologies with in situ degradation tests, enabling a “smart search” strategy in a multi-parameter space. We took both calculations and experimental data into our machine-learning decision-making step, which have led to a significant acceleration in the search process. Validation on new materials are further achieved by an employment of the synchrotron-based high-throughput XRD measurement, where the degradation profiles are directly correlated to the underlying structural changes. This work sheds light on combining theory, machine-learning and high-throughput experimentation to accelerate the development of novel solar materials.
3:40 PM Break
4:00 PM Invited
Autonomous Systems for Alloy Design: Towards Robust Closed-loop Alloy Deposition and Characterization: Brian DeCost1; 1National Institute of Standards and Technology
Autonomous research systems continually learn by adaptively planning and executing campaigns of physical and/or in silico experiments to achieve a scientific or engineering goal without direct human intervention. This emerging research area presents new opportunities to accelerate materials synthesis, evaluation, and hence discovery and design. General autonomous science systems face several challenges: learning to reliably synthesize materials, mapping material specification and processing to structure and properties, incorporating offline data streams, and incorporating prior theoretical and data-driven knowledge. As the materials community surmounts these challenges, closed-loop automated materials synthesis and characterization platforms offer much more than a means of engineering materials properties and performance through black-box optimization algorithms: they offer the potential to develop and deploy new algorithms for generating and testing scientific hypotheses. I will present two exemplar autonomous systems for alloy design that are being developed at NIST, focusing on technical and methodological aspects of building and deploying robust closed-loop synthesis and characterization platforms. The first is an autonomous X-ray diffraction system that performs active cluster analysis to efficiently map composition-temperature phase diagrams using composition spread thin films. The second is an autonomous scanning droplet cell (ASDC) designed for on-demand alloy electrodeposition and real-time electrochemical characterization for investigating the corrosion-resistance properties of multicomponent alloys. Our initial studies focus on systems that are likely to form corrosion-resistant metallic glasses (MGs) and single-phase multi-principle element alloys (MPEAs)..
4:40 PM Invited
Turning Statistical Mechanics Models into Materials Design Engines: Marc Miskin1; 1University of Pennsylvania
The core tenet of statistical mechanics is that the frequency of microstates for a material system can be used to predict its macroscopic properties. What if it were possible to turn this relationship around and use it directly for materials design? That is, instead of predicting macroscopic properties, could we engineer them by exploiting the rich information encoded in micro-states and their fluctuations? In this talk, I present a new approach that can be used to transform a statistical physics model that describes a material into a materials design algorithm that tailors it. Because the resulting algorithm is built with a physical model as its foundation, it inherits the ability to exploit micro-state information in guiding an optimization. I’ll show this extra information leads to benefits over black-box optimization methods in terms runtime, efficiency, and robustness. In particular, I’ll show examples of material optimization with this new approach, including optimal self-assembly, non-equilibrium optimization, and a real-world application on the directed self-assembly of diblock copolymers.