National Academies Press: OpenBook
« Previous: 2 Keynote Addresses
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

3

Materials Design

Lourdes Salamanca-Riba, University of Maryland, introduced three speakers who were invited to address materials design: Bill Mahoney, ASM International; Chris Wolverton, Northwestern University; and Steven Arnold, National Aeronautics and Space Administration (NASA). Tresa Pollock, University of California, Santa Barbara, moderated a short Q&A following the presentations, and then introduced a panel discussion for deeper discussion of the topic.

DATA ANALYTICS IN THE ASM COMMUNITY: BRIDGING MATERIALS SCIENCE AND DATA SCIENCE

Mahoney shared ASM’s strategies to retain relevance in 21st-century materials market development, collaborate with strategic partners, and address market challenges. By bridging materials science and data science, Mahoney sees an important role for ASM in creating “Materials 4.0,” which he considers the backbone for concepts of “Industry 4.0” or “Manufacturing 4.0.”

Retaining Relevance as a Leader in the Materials Science Marketplace

After a rocky period, ASM is in a phase of renewal with a focus on taking a leadership position in the materials science marketplace. With growing membership and revenue, Mahoney emphasized that ASM today has the patience, persistence, and resources to make it all possible. ASM is also increasing investment in its Materials Education Foundation, which encourages students to pursue careers

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

in materials, science, and engineering. Mahoney noted that the majority of ASM members come from industry, which influences how the organization approaches its mission.

ASM operates in several major areas, including Industry 4.0 and the Materials Genome Initiative (MGI), which share common underlying technologies, such as cloud computing and customization, in which ASM is fast developing expertise. Historically, the path from design to manufacturing to materials has been a sequential process, but in today’s environment, with the sharing of outputs and information, those processes can happen concurrently and in a circular way, which can create both opportunities and threats.

A Digital Transformation

ASM has recently invested heavily in new technological competencies, including a complete digital transformation of the entire organization, the reengineering of its legacy data and content to fully discoverable and shareable data, and new content and data creation from both ASM and third parties. ASM is also adding data management and services to content management, creating a new, integrated e-commerce backbone and using cloud-based Granta MI to support a new Material Solutions Network.

In addition, ASM is pursuing the private labeling and reselling of the Materials Platform for Data Science (MPDS) database; establishing the Center for Materials Processing Data (CMPD), a collaboration with universities and industry that focuses on transient data; and putting the MGI Toolkit in the cloud, where it will join ASM’s Software as a Service (SaaS) Ecosystem, an integrated computational materials engineering (ICME) model with data and tools that enables customers to conveniently find value in the Data Lake, a secure repository of all of ASM’s data that enables data and metadata experiments and analysis (Figure 3.1). The organization is also working with the Department of Defense (DoD) and Department of Energy (DOE) to create redeployable data management service offerings, as well as creating avenues for third parties to take advantage of these offerings.

The goal of this multipronged digital transformation is to make ASM’s data discoverable, searchable, accessible, and interoperable, Mahoney said. The organization will also continue to aggregate, curate, validate, and disseminate materials information, competing through volume, variety, verification, and velocity. ASM’s information and data services strategy will work with multiple participants, from apex partners to midrange and foundational marketplace levels, fulfilling its nonprofit mission of broadly disseminating materials-related technologies.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Image
FIGURE 3.1 ASM International’s Software as a Service (SaaS) Ecosystem, an ICME model with data and tools that enables customers to conveniently find value in the Data Lake, a secure repository of all of ASM’s data that enables data and metadata experiments and analysis. SOURCE: Bill Mahoney, ASM International, presentation to the workshop.

Relationship with Other Initiatives

ASM is also part of the working group for the concept evolution for International Organization for Standardization (ISO) 10303, the manufacturing interoperability standard that relies heavily on materials information. The team is exploring the exciting possibility that the standard could also be used to house and disseminate materials information, building a bridge from the materials community to the standards community.

This work is happening in parallel with NASA’s Vision 2040, an initiative closely aligned with materials science and the ICME paradigm. In addition, Mahoney said that ASM’s materials solutions information and data services will be well placed to slot into the National Institute of Standards and Technology (NIST) network-centric, computer-aided product realization and manufacturing environment.

Addressing Market Challenges

Before closing, Mahoney outlined key challenges including the sheer amount of data; the slow-moving and fragmented state of the materials data market; the

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

complexity of the technical and functional requirements, including schemas, formats, gateways, and translation requirements; and the lack of “connective tissue” in the marketplace. He stressed his view that trusted, effective consortia can push future market development, and said that ASM is actively pursuing and supporting consortia work to tackle these challenges.

Q&A

Apurva Mehta, Stanford University, asked how it was possible for open data and model sharing to coexist with business models. Mahoney replied that open access is a threat to some business models, but ASM is betting that users will be willing to pay for convenient, affordable services. He added that ASM is focused on protecting its resources through intellectual property (IP) protections and anti-piracy efforts.

In response to a question by Carlos Levi, University of California, Santa Barbara, Mahoney outlined ASM’s efforts in the education sphere. First, the effort to get the MGI Toolkit into the cloud enables universities to access these data more easily. Second, ASM camps are spreading competency in computerized interaction with materials science, complementing the experimental and physical aspects that the field has traditionally stressed. ASM believes that balancing physical and experimental methods with digital tools can also teach essential critical thinking skills.

Surya Kalidindi, Georgia Institute of Technology, asked if ASM is equipped to offer the level of customization that users want. Mahoney stressed that although customization can be daunting, ASM will listen to its customers, respond quickly, and take the lead where possible.

Prompted by Pollock, Mahoney outlined a vision for what ASM will look like in 10 years. He noted that ASM would like to be known as the Materials 4.0 platform that supports Industry 4.0. To do that, he said it must offer data in a variety of forms and formats, convenient and affordable subscriptions, and better overall data access.

USING AI TO DISCOVER AND DESIGN MATERIALS

Wolverton discussed key challenges in materials informatics, what is being done to overcome them, and areas for future study. He described four main components of the materials informatics workflow: data collection, data processing, representation of materials, and learning about properties from the results (Figure 3.2). Machine learning (ML) algorithms come into play in the last step, when they are applied to search for correlations. While these ML algorithms have garnered a great deal of attention, state-of-the-art open source ML tools are widely available, and Wolverton argued that the other steps—data collection, data processing, and materials representation—currently present the biggest hurdles.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Image
FIGURE 3.2 The materials informatics workflow involves four main components: data collection, data processing, representation of materials, and learning about properties from the results. SOURCE: Chris Wolverton, Northwestern University, presentation to the workshop, from Logan Ward, 2017, “Machine Learning for Materials Discovery and Design,” doctoral dissertation, Northwestern University, used with permission.

Data Challenges

To improve access to materials science data, Northwestern University has been assembling the Open Quantum Materials Database (OQMD), a large, open, density functional theory (DFT) database of roughly 50,000 experimentally known and 500,000 hypothetical inorganic crystalline compounds.1 The data set enables automatic computation of phase stability with respect to an arbitrary number of components.

For data to be populated into the OQMD, it needs crystal structure information to act as inputs. This information is typically derived from experiments, and makes up the 50,000 known structures. The 500,000 hypothetical structures come from prototypical crystal structure types, which are decorated with various combinations of elements from the periodic table. While many of these hypothetical structures result in high-energy, unstable compounds, Wolverton noted that even calculations that result in obvious “junk” are important because they teach the ML model the difference between reasonable and unreasonable chemistries.

It is also possible to perform high-throughput screening on the data itself, without using ML models, by searching data for materials with specific properties of interest. While this can be a powerful strategy, Wolverton said, determining a screening strategy requires some creativity. For example, thermoelectric properties can be found by searching for compounds with particular, unusual band structures, performing DFT interrogations, and screening out negative results.2

___________________

1 See http://www.oqmd.org.

2 V.I. Hegde, M. Aykol, S. Kirklin, and C. Wolverton, 2018, “The Phase Diagram of All Inorganic Materials,” arXiv:1808.10869.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

In another informatics approach, Wolverton adapted aspects of network theory to materials science, choosing the ground state convex hull of the entire OQMD as a test network. This materials network consists of about 21,000 stable compounds (nodes), which are connected via approximately 41,000,000 tie-lines (edges), which connect compounds that have stable two-phase equilibria. Wolverton and colleagues examined the connectivity of this network, enabling novel questions about compound connections, stability, nobility, and reactivity of materials.3 Another illustration of the application of network theory to materials discovery involved creation of a time-stamped convex hull network (based on the date of the first report of a compound’s discovery), allowing one to make predictions about how feasible it will be to synthesize certain materials in the future.4

Representation Challenges

The representation, a set of quantitative attributes that describe a material, is key to effective use of ML in materials science, Wolverton said. Representations must be as complete as possible, efficient to compute, accurate enough to capture important physical effects, and diverse enough to treat many possible properties. Many examples of materials representations in the literature are available, but generally these are property- and problem-specific; however, a problem-independent, general representation is most useful.

Wolverton described the use of ML in creating two types of representations: (1) those based solely on attributes determined from the composition, and (2) those that additionally include attributes determined from crystal structure information. ML models can learn from the simpler composition-based representations if enough examples of composition-property combinations are available. In many cases, this process has yielded remarkably accurate results thanks to the availability of large data sets.

These composition-only representations are also effective for predictive exploration of new materials, because an ML model can often interrogate a composition a million times faster than DFT calculations, allowing one to scan millions or even billions of compositions quickly to search for candidate materials with desired properties.5

___________________

3 M. Aykol, V. Hegde, L. Hung, S. Suram, P. Herring, C. Wolverton, and J.S. Hummelshøj, 2019, Network analysis of synthesizable materials discovery, Nature Communications 10:2018, https://doi.org/10.1038/s41467-019-10030-5.

4 E.B. Isaacs and C. Wolverton, 2018, Inverse band structure design via materials database screening: Application to square planar thermoelectrics, Chemistry of Materials 30(5):1540-1546, https://doi.org/10.1021/acs.chemmater.7b04496.

5 L. Ward, A. Agrawal, A.N. Choudhary, and C.M. Wolverton, 2016, A general-purpose machine learning framework for predicting properties of inorganic materials, Computational Materials 2:16208, https://doi.org/10.1038/npjcompumats.2016.28.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

Wolverton offered an example in which ML models were trained on experimental data available in the literature on metallic glasses. These ML models and high-throughput experiments were used in tandem to predict, synthesize, and ultimately discover novel materials with metallic glass forming ability, a property that is extremely difficult and computationally prohibitive to predict from DFT calculations alone.6

In closing, Wolverton noted that adding crystal structure information to a material representation can greatly improve the accuracy of ML models, and makes the model sensitive to polymorphic information (different crystal structures at the same composition). However, the use of these crystal structure-based representations requires knowledge of structure in order to evaluate the model, and hence can limit the utility of these approaches.

Q&A

Susan Sinnott, Pennsylvania State University, noted that the reactive force-field community has long struggled with a tension between reproducing material properties while not requiring overwhelmingly intensive computations—a tension similar to what Wolverton described in the context of his materials informatics workflow. Wolverton agreed that the two fields both struggle with representation, which will require creativity to overcome and may benefit from some convergence.

Gareth Conduit, Intellegens, asked whether manufacturing processing variables should be the basis set to describe materials. Wolverton answered that, for a specific property model, assessment and agreement on exactly which microstructural variables are important would be a step forward. In response to a question by Robert Hull, Rensselaer Polytechnic Institute, Wolverton added that outliers in large data sets (such as the OQMD) can be helpful in finding errors both in the data and in the high-throughput process and the data workflows.

Kalidindi agreed with Wolverton that representation is the most relevant problem for the materials science community, but wondered if Wolverton’s approaches use structure more than processing. Wolverton agreed that there is an implicit connection between elemental properties and structure, but that the composition-only ML models do not explicitly contain structural information. Using ML to predict crystal structures would be an interesting avenue to explore, he continued, because the current tools for DFT crystal structure prediction are computationally expensive and too slow for high-throughput approaches.

___________________

6 F. Ren, L. Ward, T. Williams, K.J. Laws, C. Wolverton, J. Hattrick-Simpers, and A. Mehta, 2018, Accelerated discovery of metallic glasses through iteration of machine learning and high-throughput experiments, Science Advances 4(4):eaaq1566, http://doi.org/10.1126/sciadv.aaq1566.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

VISION 2040 STUDY: NASA’S IMPLEMENTATION ACTIVITIES

Arnold gave an overview of NASA’s Vision 2040 initiative, its implementation, and current projects intended to populate its ecosystem. He emphasized that Vision 2040 is not exclusive to NASA; rather, it is an effort for the broader community, guided by hundreds of experts from academia, government, and industry who participated in its development. Realizing the vision will require collaboration among industry, research laboratories, and multiple federal agencies.

Overarching Goals

The Vision 2040 initiative is part of a broader transition to a design future rooted in ICME—a transition that promises to bring the materials community from its past paradigm of designing with the material to a future model of designing the material itself (Figure 3.3). Therefore, Vision 2040 defines a roadmap for marrying these two paradigms into a single integrated, multiscale modeling and simulation of materials and systems enabled by a cyber-physical-social ecosystem that accelerates model-based design.

Image
FIGURE 3.3 Integrated computational materials engineering (ICME) is an approach to the design of products and the materials that comprise them by linking experimentally validated materials models at multiple length scales. SOURCE: Steven Arnold, National Aeronautics and Space Administration, presentation to the workshop.
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

Arnold stressed that Vision 2040 aims to inspire true digital transformation and not merely digitization of the enterprise. Aerospace is the end result because NASA is the primary funder, but all of the work can be extended into the entire materials ecosystem. The roadmap outlines a path toward a concurrent design environment, applicable throughout the entire supply chain, which can create fit-for-purpose materials throughout the entire system. It seeks to replace today’s paradigm in which design is disconnected, tools are domain-specific, and materials properties are based on empirical testing with a more integrated way of working that seamlessly joins stages of the product development life cycle, tools are usable across the community, and material properties are determined virtually with a heavy emphasis on simulation rather than physical testing.

The Vision 2040 Ecosystem

Vision 2040 identifies nine key element domains that ultimately should be interwoven: models and methodologies, multiscale measurement and characterization tools and methods, optimization and optimization methodologies, decision making and uncertainty quantification and management, verification and validation, data and information and visualization, workflows and collaboration frameworks, education and training, and computational infrastructure (Figure 3.4).

Image
FIGURE 3.4 Elements of the materials ecosystem that will be interwoven as part of Vision 2040. SOURCE: Steven Arnold, National Aeronautics and Space Administration, presentation to the workshop.
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

Although the report contains 117 gaps, one critical gap was identified for each element domain, as well as the actions, time frame, and cost needed to close that gap and the desired end state characteristics to create practical, useful knowledge for industry. Ten cross-cutting streams were identified to help organize these gaps and recommend actions across key elements. The data analytics and visualization stream, for example, identified key limitations with regard to representing time-dependent data, representing translucency among multiple layers of data, and using AI/ML to improve scalability.

Vision 2040 also outlines 11 major recommendations for NASA, such as to create an interagency coordinating body; facilitate partnerships with other agencies and organizations to develop an interoperability framework; identify and pursue benchmark materials systems and applications; and create, maintain, and disseminate “gold-standard” data sets. A key recommendation is to stimulate widespread cultural change, which, Arnold noted, can present an even greater barrier than creating the necessary technology. Small Business Innovation Research (SBIR) solicitations have also been restructured to align with Vision 2040.

Implementation

In terms of immediate areas of emphasis in materials and systems design, Arnold highlighted how NASA is approaching three areas of focus out of nine Multidisciplinary Engineering Challenges (MECs) outlined in the vision: mitigating damage in high-temperature engines, designing unique alloys for aerostructures, and aircraft electrification. Reaching these goals will require new methods, visualizations, and applications, some of which are ML-based. NASA’s Revolutionary Tools and Methods (RTM) “swim lanes” framework outlines how integrating from first principles to macro materials design—and data-driven and physics-based approaches—will enable rapid materials discovery and lead to design and certification of advanced manufacturing-based materials and structures.

Arnold discussed progress toward key elements of Vision 2040 including models and mechanisms, computational infrastructure, and optimization. In terms of models and mechanisms, for example, the NASA Multiscale Analysis Tool (NASMAT) is being developed as a state-of-the-art, thread-safe, multiscale modeling program. Other areas of development include new tools to build a material designer infrastructure to support design-rule discovery, screening, and multiscale/multiphysics assembly, with target applications in coatings, metallic alloys, and battery materials. Test results for nickel-based superalloys are promising, and NASA has found that employing ML can accelerate this work, providing not only results but also explanations to expand understanding.

Virtual testing is another important area, and one that can lead to significant cost savings, Arnold said. In this context, ML is being used to develop surrogate

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

models to dramatically speed up those virtual tests and determine what types of real tests need to be performed for maximum information extraction. NASA is also developing schemas for enabling both real and virtual data to reside in the same information management system, enhancing data management. As a result, NASA is building model, microstructure, and software tools tables along with script-driven, generic computational frameworks to support ICME and increase testing speed with zero data loss.

Q&A

Wolverton asked if high energy density batteries were being pursued to electrify aircraft. Arnold replied that they are, and noted that the limiting factor right now is weight. NASA is fostering collaborations and exploring other materials besides electrolytes, such as magnets, metals, and composites, to push progress in this area as well.

June Lau, NIST, asked if the SBIR program, which encourages IP and data ownership by businesses, can coexist with collaboration and openness. Arnold agreed that some tension exists, but he believes it is possible, although difficult, to find a balance between widespread participation and ensuring adequate protections for information as appropriate.

PANEL DISCUSSION ON MATERIALS DESIGN

Pollock introduced the speakers for the workshop’s first panel discussion: John Mauro, Pennsylvania State University; Brian Storey, Olin College and Toyota Research Institute; and Dane Morgan, University of Wisconsin, Madison. Salamanca-Riba moderated a Q&A period following the panelists’ introductory remarks.

John Mauro, Pennsylvania State University

Mauro is a glass scientist with both industry and academic experience. His work in glass science and computer science merges data informatics with materials design. He discussed the process of designing materials with desired properties and some challenges for data science.

For new materials to truly change the world, they must be manufacturable, and so new materials design must balance product and manufacturing attributes, all of which are changing simultaneously with composition. Models are a great way to accelerate this process, and multiple types of models—not just ML but also data-driven, statistical, continuum, and analytical models, to name a few—used in combination lead to the best insights and predictions.

For each desired property, models turn the design process into a mathematical optimization exercise, where one property is maximized by adjusting the

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

composition, while bearing in mind the constraints of the other properties needed for performance and manufacturability. In his work at Corning, Mauro used this combination of physics-based and empirical-based modeling to create a model-predicted glass, Gorilla Glass, with optimum composition that dramatically lowered costs and improved performance, accelerating the path to full-scale manufacturing.

Mauro identified several key challenges for data science in materials, including data integrity, data security, and creating and maintaining dynamically updated models. He expressed confidence that the technical challenges can be overcome but suggested that changing institutional culture presents a significant challenge to widespread adoption of data science methods in materials design.

Brian Storey, Olin College and Toyota Research Institute

The Toyota Research Institute (TRI) was founded in 2016 with a $1 billion investment from Toyota to advance technology in three areas: Accelerated Materials Design and Discovery (AMDD), automated driving, and robotics. The key goal of AMDD, Storey’s area, is to accelerate emissions-free technology by advancing battery and fuel cell technology.

AMDD has a small internal team and a research consortium with $10 million invested across 12 universities and national laboratories. Each project is a collaboration, creating a community around data and tools. Current projects include BEEP, to accelerate research and development for batteries at the systems scale by reading and interpreting battery data; and ACE, to improve catalyst performance by automating high-throughput experiments and data analysis.

AMDD is also building online infrastructure and modules for collaboration, offering real-time data sharing, model adjusting, and research automation. Many of its research tools are available online, including BEEP, ACE, Network, PROPNET, and MATSCHOLAR, which performs natural language processing on materials abstracts.

Dane Morgan, University of Wisconsin, Madison

Morgan’s laboratory focuses on integrating atomistic foundations with higher level theories, including using tools from data science, to make predictions for materials applications. ML tools are having a big impact in materials science in many areas, including by making predictions about new data in order to guide future experiments or enhance data sets, replacing physical models with more efficient correlative data-centric models, and interpreting and extracting data from images and text.

While using ML in materials science and engineering is not new, Morgan said the explosion of new tools and capabilities brings new opportunities to transform

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

materials science. There are also challenges, including creating accessible infrastructure for disseminating models; continuing software development; collecting input from and engaging with data scientists; creating a closed loop design to iteratively develop materials; offering open, accessible data; and exploring more “hard” problems related to innovation. An exciting and potentially transformative aspect of ML and AI would be to use them to help with creative exploration and understanding, not just the more well-defined and constrained parts of design and discovery, Morgan suggested.

OPEN DISCUSSION

Following the panelists’ remarks, participants discussed specific uses of ML for materials design, unique aspects of materials design as compared to other domains, cultural issues in this space, and opportunities for collaboration.

ML Applications in Materials Design

Manoj Kolel-Veetil, U.S. Naval Research Laboratory, asked if ML can predict what will happen under particular conditions for one specific property, even if the result is something completely novel and unexpected. Morgan replied that ML is well suited for such a situation, assuming the problem is well defined and there is enough data on the property, chemistries, and related structures to train the model in such a way that the new behavior is within the domain of the model. However, it is unclear if results could surpass known property limits, because extrapolation beyond training data values is often unreliable.

Salamanca-Riba asked if, in addition to ML, it was possible to use robotics and automation in materials design. Storey answered that it was an exciting idea, but cautioned that not everything can or should be roboticized. Automation, he said, should be used only to address real problems or bottlenecks, and only if it makes sense to do so. Another challenge, he continued, is the difficulty moving from discovery of new materials to a useful market application. Mauro added that robotics can be a very brute-force approach, and, like blind high-throughput discovery tests, does not take advantage of what is already known about chemistry and reactions. A combination of ab initio experiments, ML models, and high-throughput discovery would be more successful for creating useful materials, he concluded.

Levi asked if ML can provide insights on structural changes. Mauro replied that ML models can be based on different sets of predictors, including chemical structures such as individual oxides. ML models could also be tweaked by including chemical structures in the predictors, and the result would then include useful information about the predictors themselves.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

Wolverton argued that the use of ML to find fundamental physics or physical laws in data sets is very difficult. He cautioned that part of the difficulty is owing to the fact that although ML is good at finding correlations, correlation is distinct from, but easily mistaken for, causation. Morgan noted that, while this is in general true, there is nevertheless an active subfield of ML that seeks to extract physics equations from data. Arnold stressed the importance of application-driven design; in his view, ML should be used to combine material and structural aspects in order to reduce the time spent on nonpromising materials and account for “lurking failure modes.”

James Warren, NIST, asked what machines can do at a multiscale level, starting from an equation. Mauro replied that researchers are still learning about each scale—such as electronic, structural, microstructural—and a multiscale approach that incorporates physics and ML represents an excellent candidate for a Grand Challenge for the field.

Another participant asked about using ML for composite materials. Storey replied that for catalysts and batteries, surface properties are poorly understood, and right now it is unclear that ML alone will help. In general, however, moving from bulk structure introduces more and more complexity, and likely requires a data-driven coupling of data and simulation. Could learning about surface effects create a transition to multiscale composites? Storey replied that the answer is not clear at this point. Adding that ML is difficult without a lot of data, Morgan suggested that the answer may depend on how the specific problem is defined. ML and other tools can make a short-term impact here, but tools that automate the workflow can have a large impact, even if they are not technically “discovering science.”

Another participant asked about ethical considerations when using ML. Mauro replied that a machine can learn ethics if there are definitions of ethical and unethical behavior and enough examples for machines to learn from. Deciding what a machine should be allowed to do is a different, much thornier question, however. Morgan agreed, adding that ethics is a big issue with self-driving cars, genomics, medicine, and other areas, although it has not yet presented major concerns in the realm of materials science.

Unique Aspects of Materials Design

In reply to a question from Kalidindi, Mauro noted that unlike engineering, ML for materials design does not have formal methodologies in place. He added that even in laboratories with formal methodologies, different scientists have always approached materials design from different viewpoints. In some sense, he continued, it may be better to not have a formal methodology, because having different viewpoints can be a strength in that it leads to unexpected insights and discoveries.

When asked by Kalidindi if formalism could speed up materials innovation,

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

Mauro stated that ensuring scientific rigor and best practices would be preferable to strict methodologies. Storey added that formal processes fail to address the cultural and personal aspects that are intrinsic to design, which relies on adaptable tools and processes to be effective.

Pollock asked if an optimization problem is also at play. Mauro replied that materials designers can either seek materials that meet certain requirements—recognizing that in practice, the target frequently changes, resulting in a long, frustrating, and decidedly nonoptimized research cycle—or they can build modeling tools based on experiments, physics, and ML that can adapt to changing requirements more quickly—an approach that, in his view, capitalizes on the true value ML offers for materials design.

Mehta asked how materials design balances the differences between science and engineering, and especially the tension between structure and properties. Mauro agreed that the tension is real. From an engineering perspective, structures are useful only if they add to the chemistry and processing data, improving predictions. ML can be valuable in making predictions from composition to structure, including microstructure, and combining those with data to get to final properties. It is important to verify the models’ output to ensure that the result is an optimized material with real-world impact.

Mehta then noted that some problems are not solvable by humans, and only ML can create the fundamental change needed to solve them. Mauro suggested that these problems need humans and computers, working together. If only known data are relied on, accurate predictions are not possible, but known data plus ML provides the best opportunity to extrapolate into the unknown.

Noting the large disconnect between materials science and the finite element community, another participant suggested that focusing on the unifier, thermodynamics, which is at the heart of all continuum mechanics, would be beneficial. Storey agreed, but generalized the unifier to nonequilibrium thermodynamics.

Cultural Issues and Collaboration

Warren asked the panelists to talk about the cultural problems they have encountered using ML in materials design. Mauro answered that the development process that resulted in Gorilla Glass was effective partly because the team was so small, and attempts to replicate its success more broadly came up against two serious cultural issues. First, there was a large disconnect between theoreticians/modelers and experimentalists. Second, there was tension between scientists who wanted solo credit and others, usually younger, who were more open to trying new approaches, sharing data, and working in teams. He recommended that companies restructure their reward systems to recognize teams, not just individuals.

Haydn Wadley, University of Virginia, asked if the materials and application

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×

manufacturers could merge their expertise and create a more holistic materials design and discovery environment. Storey noted that TRI recognizes the value of partnerships to solve these problems. Morgan added that it is possible to work together and create an open innovation infrastructure. Having clear targets and publicly available data created improved fuel cell catalysts, for example. Mauro added that end users, perhaps from a lack of understanding of materials trade-offs, often do not know what they really need. One of the materials industry’s biggest challenges is to build rapport with customers that will ultimately lead to better designed products, although companies are sometimes reluctant to share confidential information. They do not have a culture of openness about their data, which complicates the process, he added.

Pollock asked the panelists what collaboration platforms or tools they found most helpful or unhelpful. Storey answered that these problems are so challenging that they require collaborative engagement of a broad community. Also, tools cannot be developed in a vacuum—they need to solve real problems. On the other hand, early progress often has to be demonstrated before people are willing to be engaged. Mauro shared several examples of what not to do: build tools in silos; build tools without sharing them or providing training; and refuse to use community tools. In addition, he noted that competition can make companies reluctant to engage with larger efforts. Morgan added that many young researchers he meets are very open to working in a highly collaborative, sharing environment, which he sees as a promising direction for the field.

John Gardner, NASA, asked the panelists how the portrayal of AI/ML in popular media affects their work. Mauro replied that it is generally not helpful to have expectations that exceed reality, and for popular culture to generate fear around AI/ML. Morgan countered that media coverage does inspire people to engage with ML and generates overall interest, so there are pros and cons.

Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 13
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 14
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 15
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 16
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 17
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 18
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 19
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 20
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 21
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 22
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 23
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 24
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 25
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 26
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 27
Suggested Citation:"3 Materials Design." National Academies of Sciences, Engineering, and Medicine. 2021. Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop. Washington, DC: The National Academies Press. doi: 10.17226/25628.
×
Page 28
Next: 4 Data Curation »
Data Analytics and What It Means to the Materials Community: Proceedings of a Workshop Get This Book
×
Buy Paperback | $40.00 Buy Ebook | $32.99
MyNAP members save 10% online.
Login or Register to save!
Download Free PDF

Emerging techniques in data analytics, including machine learning and artificial intelligence, offer exciting opportunities for advancing scientific discovery and innovation in materials science. Vast repositories of experimental data and sophisticated simulations are being utilized to predict material properties, design and test new compositions, and accelerate nearly every facet of traditional materials science. How can the materials science community take advantage of these opportunities while avoiding potential pitfalls? What roadblocks may impede progress in the coming years, and how might they be addressed?

To explore these issues, the Workshop on Data Analytics and What It Means to the Materials Community was organized as part of a workshop series on Defense Materials, Manufacturing, and Its Infrastructure. Hosted by the National Academies of Sciences, Engineering, and Medicine, the 2-day workshop was organized around three main topics: materials design, data curation, and emerging applications. Speakers identified promising data analytics tools and their achievements to date, as well as key challenges related to dealing with sparse data and filling data gaps; decisions around data storage, retention, and sharing; and the need to access, combine, and use data from disparate sources. Participants discussed the complementary roles of simulation and experimentation and explored the many opportunities for data informatics to increase the efficiency of materials discovery, design, and testing by reducing the amount of experimentation required. With an eye toward the ultimate goal of enabling applications, attendees considered how to ensure that the benefits of data analytics tools carry through the entire materials development process, from exploration to validation, manufacturing, and use. This publication summarizes the presentations and discussion of the workshop.

  1. ×

    Welcome to OpenBook!

    You're looking at OpenBook, NAP.edu's online reading room since 1999. Based on feedback from you, our users, we've made some improvements that make it easier than ever to read thousands of publications on our website.

    Do you want to take a quick tour of the OpenBook's features?

    No Thanks Take a Tour »
  2. ×

    Show this book's table of contents, where you can jump to any chapter by name.

    « Back Next »
  3. ×

    ...or use these buttons to go back to the previous chapter or skip to the next one.

    « Back Next »
  4. ×

    Jump up to the previous page or down to the next one. Also, you can type in a page number and press Enter to go directly to that page in the book.

    « Back Next »
  5. ×

    Switch between the Original Pages, where you can read the report as it appeared in print, and Text Pages for the web version, where you can highlight and search the text.

    « Back Next »
  6. ×

    To search the entire text of this book, type in your search term here and press Enter.

    « Back Next »
  7. ×

    Share a link to this book page on your preferred social network or via email.

    « Back Next »
  8. ×

    View our suggested citation for this chapter.

    « Back Next »
  9. ×

    Ready to take your reading offline? Click here to buy this book in print or download it as a free PDF, if available.

    « Back Next »
Stay Connected!