Symposium DS01-Accelerating Experimental Materials Research with Machine Learning

In almost all areas of materials research, reliable knowledge can only be gained by performing experiments. In these areas, the pace at which knowledge gained is highly dependent upon both the rate at which experiments can be completed and the choice of which experimental conditions to probe. Recently, machine learning and automation have become major players in both of these areas by accelerating the pace of experiments and choosing experiments in a manner that ensures the generation of new knowledge. While these approaches have already provided breakthroughs in fields ranging from nanomaterial growth, electronic property selection, and mechanical structure design, they have also unified a community of researchers through the uncovering of new challenges unique to these novel human-machine partnerships. While this community includes both active learning systems in which experiments are chosen and interpreted by machine learning, and autonomous research systems in which experiments are also performed without human intervention, all systems have to address challenges regarding structuring the machine learning process, providing prior knowledge, incorporating uncertainty, and fruitfully leveraging the human-machine partnership. The symposium will highlight achievements and challenges from these fields of active and autonomous research ranging from the presentation of new materials discoveries made using such platforms to fundamental innovations in the development of machine-learning guided experiments.

Topics will include:

Materials discoveries made using autonomous research systems
Materials discoveries made using machine learning guided experiments (active learning)
Comparisons of conventional high-throughput experimentation and active learning
Benchmarking methods for quantifying efficacy of active learning methods
Generality vs. specificity in terms of experimental platform development, including hardware, software, and ontologies
Virtues and limitations of Bayesian optimization and the role of decision-making policies
When is property vs knowledge maximization a false dichotomy and when is it a necessity
Uncertainty quantification and propagation for machine learning modeling of physical process
Accommodating modeling systematic uncertainty
Automated physical modeling and scientific learning
Automatable infrastructure including hardware/software and distributed systems
Human-Machine partnering in Materials Research including visualization tools for active learning
Limitations of Gaussian Processes in describing materials systems
Transfer learning, multiple-information source optimization, and contributions from simulation

Invited Speakers:

Milad Abolhasani (North Carolina State University, USA)
Alan Aspuru-Guzik (University of Toronto, Canada)
Christoph J. Brabec (University of Erlangen, Germany)
Tonio Buonassisi (Massachusetts Institute of Technology, USA)
John Gregoire (California Institute of Technology, USA)
Jason Hein (The University of British Columbia, Canada)
Kedar Hippalgaonkar (Agency for Science, Technology and Research, Singapore)
Amanda Krause (University of Florida, USA)
Benji Maruyama (Air Force Research Laboratory, USA)
Elsa Olivetti (Massachusetts Institute of Technology, USA)
Kristin Persson (University of California, Berkeley, USA)
Kris Reyes (University at Buffalo, The State University of New York, USA)
Helge Stein (Karlsruhe Institute of Technology, Germany)
Ichiro Takeuchi (University of Maryland, USA)