6
Discussion
Haydn Wadley, University of Virginia, wrapped up the workshop by posing a series of broader questions on the roles of data and machine learning (ML)/artificial intelligence (AI) in advancing materials design and discovery.
HOW WILL DATA, AI, AND DEEP LEARNING TRANSFORM MATERIALS DESIGN?
Apurva Mehta, Stanford University, expressed his view that materials science is changing and posited that AI will have an increasingly larger role in those changes. The nature of the data, data analytics tools, and overall data and computation landscape that develops will depend on whether funding and policies come from government or industry. It may be helpful, he continued, to define a concrete vision of that landscape to motivate work toward achieving it.
James Warren, National Institute of Standards and Technology (NIST), agreed, adding that he believes that materials science going forward will be driven primarily by a model of small groups working together; establishing a larger, more integrated system would require massive government funding, which he sees as unlikely. Warren added that the most likely scenario is that automation will slowly ramp up, first in fields where it is more tractable, such as small-molecule chemistry, and then eventually in materials science. Despite its complexities, AI will change materials science, and probably sooner than we think, he concluded.
Susan Sinnott, Pennsylvania State University, asked if the goal was for materials scientists to catch up to computational chemists and drug designers, who have
more experience with ML, or if materials work faces different problems. Manoj Kolel-Veetil, Naval Research Laboratory, suggested that materials scientists could learn from pharmaceutical work—for example, from the model of personalized drug design.
Surya Kalidindi, Georgia Institute of Technology, reiterated that design is a separate discipline with its own tools, and is very different from discovery. True materials design may require collaboration with professional designers who can incorporate materials data into their workflows. Gareth Conduit, Intellegens, noted that an important goal for AI in materials is concurrent design, where ML can be used to design a material for exact parameters and applications instantaneously, rather than relying on databases of existing materials. He noted that in this respect materials design is different from chemistry applications, which require greater involvement from human researchers.
IS THERE AN ECONOMIC MODEL TO SUSTAIN DATA ANALYTICS?
June Lau, NIST, asserted that data is a currency, and posited that a sustainable model is needed that avoids dividing the materials science world into haves and have-nots. Bill Mahoney, ASM International, expressed his belief that data work is economically sustainable, pointing to the fact that the big data analytics market is worth $50 billion annually and growing fast, particularly with smaller, niche providers of independent data services. For its part, ASM is actively pursuing opportunities to create a full suite of services supporting materials data analytics.
A participant noted that he has had good experiences with niche companies building custom technologies, and suggested that NIST, which has so much data and experience, can also play an important role. Based on Intellegens’s market analysis of the value of materials discovery, Conduit said that while it is good to have successful laboratory results, success in a factory is where the real value is added, and so should be a goal for AI in materials and manufacturing.
Mehta added that open source AI and ML data analytics tools are both readily available and relatively easy to develop, and could also have a significant impact. Warren noted that academic and corporate values are frequently at odds, and while open source tools can be very effective, they do not always scale well. Similarly, he said that promoting access to tools and inspiring the next generation of tools requires benchmarks, evaluations, and an understanding of the potential for commercial applications for every tool created.
WHAT IS THE ROLE FOR GOVERNMENT?
Warren posited that the challenges in this space are so large that the government cannot do everything, but initial government investment can influence
underlying elements that support the entire ecosystem, such as infrastructure, standards, and process ordering, much like the government’s role in the development of the Internet. Tresa Pollock, University of California, Santa Barbara, added that infrastructure investment is also required to generate the much-needed data to spur progress in AI.
Conduit added that there is a need for certifying and standardizing models, in addition to materials, which could be a good role for government. Warren noted that NIST is considering benchmarks to allow others to assess and evaluate models.
WHAT ROLE WILL DATA ANALYTICS PLAY IN AN ERA OF MATERIALS BY DESIGN?
John Gardner, NASA, reiterated that designing a new material is not enough; it must also be able to be shaped into a working part. Therefore, the application and manufacturing environment will determine how much of a role data analytics plays.
Ichiro Takeuchi, University of Maryland, suggested a paradigm shift where AI could be applied directly to manufacturing, perhaps fixing materials as they are being produced. Gardner agreed that it was possible, but such a process could also introduce defects that would need correcting.
Pointing to developments in the Internet of Things (IoT), Mehta suggested that a connected loop of sensors, analytics, and decision support via Bayesian processes or other statistical analyses could benefit materials science, but would require new infrastructure. Warren agreed that this idea holds potential, although there are benefits and drawbacks. For example, IoT equipment can more easily be standardized and connected, eliminating proprietary data formats to enable better data sharing. It is also important to note, he continued, that the increasing popularity of maker spaces and democratized manufacturing can disrupt the current manufacturing landscape in unpredictable ways. Mahoney added that disaggregation of the traditional manufacturing base with new, smaller firms that are more adept at using AI can undermine the larger manufacturers.
IS THE WORKFORCE PIPELINE STRONG ENOUGH TO SERVE THE NEEDS OF THE FUTURE?
Wadley noted that more funding for education is needed, and Kalidindi added that the materials science curriculum also needs to change dramatically. Gardner suggested that a basic understanding of ML, AI, and statistics should be emphasized, and Mehta stated that materials science needs to offer a compelling vision in order to recruit the brightest students to the field in the coming decades.
This page intentionally left blank.