Perspective: Materials informatics across the product lifecycle: Selection, manufacturing, and certification

The materials development process is sometimes viewed through a lens of materials discovery by researchers, processing and qualification by those that manufacture materials, and selection and lifecycle management by those that design products requiring advanced materials to be used within them. In their 2001 report, the National Research Council writes that taking a “new consumer product from invention to widespread adoption typically takes 2 to 5 years, but doing the same for a new material may take 15 to 20 years”.¹ Important to note is that this assertion does not include the process of actually inventing a new material at the lab scale. Indeed, even after the long and risky process of invention is over, there exists a further delay before a material can be effectively scaled from grams, to tons, to kilotons and finally integrated into production. Some examples of materials timelines can be seen in Table I. In the strategic plan of the Materials Genome Initiative, connecting materials development, manufacturing, and lifecycle is identified as an important strategic goal.² 

This disconnect is well-known and many researchers and organizations have attempted to address it in different ways. In 1995, NASA published guidelines around Technology Readiness Levels meant to systematize documentation of how close a new technology is to flight readiness.⁸ Shortly thereafter, the Department of Defense (DOD) identified that in addition to technology readiness, ability to manufacture the technologies—many of which are materials enabled—is critical to the deployment of those technologies. As such, the DOD created the Manufacturing Readiness Level system.⁹ These two systems emphasize that manufacturing and technology development must go hand-in-hand to achieve real-world impact. The comparison of MRL to TRL numbering can be seen in Table II.

While the high level problem is obvious and easily stated, the systems and norms required to accelerate the process are complicated because they cross disciplines, length scales, and corporate boundaries. They require working with advanced algorithms, cutting edge database technologies, and heterogeneous data. Addressed systematically, none of these is insurmountable; but similarly, none can be addressed in isolation. If the goal of materials development is fundamentally to enable better product development, then taking the perspective of a product developer is critical. This person needs to understand how to select the optimal material for use, produce that material at scale, and know how the material will behave over its lifetime. This paper will discuss the challenges and opportunities in the selection, manufacturing, and qualification of advanced materials with attention paid to the role materials informatics can play in enabling better decision making in each of these areas to accelerate the development of physical products.

Identifying a material to fit a specific use case, requiring the satisfaction of multiple intrinsic property and compatibility constraints, is a core problem in engineering design. While significant effort is being directed at green field discovery of new materials, it is crucial not to neglect the problem of selecting optimal solutions from within the known space. Not every new application requires a radically new material, and over the past five decades a large number of alloys, composites, and polymer materials have been developed and integrated into engineering products. Engineers continue to find value in existing materials by exploiting commonalities between ostensibly disparate use cases, where it is the unique combination of properties that distinguishes the winning material, not necessarily record breaking performance in any individual category. Consider the superalloy Inconel (625), a superbly strong and corrosion-resistant alloy originally developed as a structural material for supercritical steam power plants,¹⁰ which was introduced this year into the battery contacts of the Tesla^® Model S to take advantage of its stress response at the extreme temperatures that result from resistive heating of the contact during rapid acceleration.¹¹ Or take the example of poly(ethylene-vinyl acetate), a copolymer thermoplastic, which was selected as the matrix material for the NuvaRing¹² for the same combination of mechanical properties that led Crocs™ to embrace PEVA, in the form of Croslite^® foam, as the primary structural component of their trademark footwear. Examples of cross-pollination between disciplines and product classes are surprisingly rare, but given the overlap in design considerations one wonders why it cannot be more commonplace. Exploiting commonalities in engineering requirements across domains will become only more critical in the future if continued investment in Integrated Computational Materials Engineering (ICME) modeling and informatics succeeds in accelerating the rate of materials discovery and increasing the search space for possible materials.

In the context of statistical modeling and optimization, the task of materials selection can be recast as an inverse problem: given the massive catalog of characterized materials, identify a subset of those materials that meet a set of desirable property conditions. Formally, we want to identify the subset of materials that represent as close an approximation as is available to the Pareto front,^13,14 the surface over which any incremental improvement in one material property is attended by the commensurate reduction of another. Further selection from this set can be made by constructing a scalar-valued fitness function with penalties associated with each property tradeoff (e.g., a 10% reduction in yield strength may be worth a 5% increase in corrosion resistance), but the principle challenge is in sampling from Pareto optimal materials.

When the properties of interest are directly computable from the description (“genome”) of a proposed material, as in alloy design using the ICME modeling framework, one can in principle generate candidates computationally without resorting to data mining (setting aside the need to parameterize the underlying models), converging on an estimate of the Pareto optimal set using established methods such as differential evolution or similar strategies that have been shown to work well on this class of problem.¹⁵ In many cases, however, the properties of interest, such as corrosion resistance and aging for alloyed metals,¹⁶ are either too difficult or costly to compute using ICME. The computational expense, combined with the danger of modeling approximations introducing critical prediction errors, has motivated an increase in the use of data-driven methods for statistical modeling popularly referred to as response surface modeling, or its generalization machine learning.^17–19 These methods seek to learn a representation of the data that is able to accurately generalize to new materials or formulations. Through this lens, material selection and optimization can be seen as tightly coupled problems: by modeling the properties of a large catalog of known input materials, one can rapidly screen and subsequently interpolate between previously characterized materials, providing a single platform with which to decide whether an existing material or derivative will meet a given set of engineering requirements. Proof of principle applications for this approach have already been established,^17,20 and we anticipate that this list will grow rapidly in the near future. Arguments can be, and have been,^21,22 made for the ability of this approach to generate truly novel scientific insights and guide discovery of new materials, but it is perhaps the incidental achievement of a platform that enables this kind of search that will have the most near term impact.

Producing a target material, particularly a novel material, at scale presents tremendous challenges. To select a material without an understanding of reliable processes that can be used to make it overlooks a major challenge in product development. As long ago as 1995, Thomas Eagar of the MIT Technology Review wrote “since process technology is likely to be the key to a company’s success in commercializing new materials, companies need to reevaluate the time and money they spend on developing new materials. In particular, they need to foster cooperative research in the early stages of materials development.”³ Around that time, the hypothesis was that communication among the various stages of the product lifecycle—product market, materials technology, process, and manufacturing capabilities—should help to reduce the 20 year period of development from invention to deployment.²³ Today, that gap still exists. While communication among those stages continues to be well established as leading to improved ROI and competitive advantage,²⁴ something additional has been missing in the acceleration of the deployment process. Communication, we assert, is only a first step; coordinated knowledge sharing across the materials lifecycle is a key. Informatics tools play a key role in just such integration.

Despite the perhaps obvious benefits of ensuring materials discovered or selected for use be manufacturable, there remain major challenges in the development of new materials to ensure that they are manufacturable at scale. ICME has handled this well to date, using processing techniques in models that are generally widely available leading to well-known systems taking advantage of well-known processing techniques to achieve new performance targets. A challenge remains, though, in the application of ICME-style approaches to materials systems for which deep mechanistic knowledge does not exist. Thermoelectrics, batteries, superconductors, and many other classes of advanced materials lack the clear mechanistic knowledge of their materials that the ICME-focused segments of the alloy world have. As such, small variations in processing or modifications of process from lab scale, to pilot scale, to production scale can substantially undermine or completely eliminate the benefits of using a new material at all because of reduced yield or reliability.

Outside of the alloy world, examples do exist where new materials were successfully invented and deployed to wide market adoption. A notable example is the development of NbFeB magnets in the 1980s. Research originally intending to discover a low-coercivity magnet as an alternative motor core soft magnet led to a discovery of a high-coercivity magnet when boron impurities from the melt spinning apparatus (borosilicate glass ampoules) resulted in the introduction of boron.⁵ By the end of the decade, the NdFeB rare earth magnets were commercially adopted and rapidly replacing incumbent materials. By 1990, 16.5% of the world market for magnets generally belonged to NdFeB, surpassing the incumbent SmCo magnets in market share by dollar value.²⁵ They not only had superior magnetic properties but had supply chain advantages: the supply of cobalt was being threatened by political unrest in Africa. The availability of an emerging alternative, combined with externalities on supply, led to the rapid adoption of alternative technologies. It is reasonable to conjecture, and supported by comments of those in the industry at the time, that such constraints provided stimulus necessary for the organizations within a supply chain to communicate more effectively across organizational boundaries the needs, requirements, and conditions necessary for an alternative to be adopted. Key to this was the effective conversion of data—information about a scientific breakthrough—into actionable knowledge—the price, performance, and requirements necessary for that technology to be adopted in a new material product—a critical factor in reducing the discovery to adoption time in materials research and development.

Even with a successful history of modeling materials in production, challenges loom on the horizon in parallel with those of the materials industry: systems are becoming more complex while new, less well understood methods are being called for. This combination only increases the need for analytics systems to be integrated with manufacturing processes.²⁶ Indeed, such issues have been well-known in the pharmaceutical industry for some time. Take, for example, the use of crystallization for separation and purification in drug manufacture where direct control in batch manufacturing processes can yield benefits in systems that are understood phenomenologically rather than mechanistically.²⁷ These processes, often too complicated to model from a purely theoretical basis, benefit from a grounding in empirical analysis. When a better theoretical understanding of the process is attained, it is then used to supplement the empirical model. The same is true of the materials industry. Starting with phenomenological and empirical models creates an opportunity to accelerate development of these systems rather than waiting for robust mechanistic theory to fully mature.

All of these opportunities are challenging for traditional methods of determining materials process and performance limits. How does a certification body create standards for processes with nearly infinite degrees of freedom? Perhaps additive manufacturing (AM), the industry equivalent of 3-d printing, is the best example of a leap forward in both capability and control of materials and product manufacture. Such advancements come with process tradeoffs and knowledge gaps.²⁸ As such, the immediate impact of AM is being seen in tooling for reductive manufacturing approaches. As the industry matures, it will need to overcome the challenges inherent in creating degrees of control without a full understanding of how that control impacts long-term performance under extreme conditions.²⁹ 

The use of complicated, novel material systems in performance critical structures and devices creates the need for model driven qualification techniques. Take, for example, the case of orientation controlled AM, which has been demonstrated recently.³⁰ Similar control of microstructure, composition, physical dimension, and other processing undoubtedly can enable improved performance across a variety of metrics. This control creates a level of analytical complexity that has not been seen before in manufacturing technologies creating a need for new approaches to materials certification that integrate materials data with process models and lifetime analyses. Indeed, because the problem is combinatorial in nature, the number of standards to address the possible instances of a single composition is unfathomably high. At the same time, with increased pressure to shorten product development cycles, there is no doubt that the insertion of AM processes into production lines will create pressure to more rapidly valuate new materials and structures as designers and engineers begin to take advantage of AM specific capabilities.³¹ Such a confluence—unprecedented control and volumes of data—creates real opportunities for materials and product, but only if the data can be effectively communicated and analyzed.

These complicated data generated in both the manufacturing and measurement of products using finely controlled tools demand a language with which to represent them in the form of data standards. While test certificates have existed for some time, even standardized data formats for electronic test certificates for engineering test data of materials have been identified as lacking and been called for.³² While the discussion around what these formats may look like continues, this call couples exactly with community coalescence around a similar data standards for fundamental materials data as well as processing information. By bringing the three of these together in compatible ways, both communication around data and advanced analysis enabled by data can be leveraged across the materials lifecycle. Private entities (Citrine Informatics), non-profit societies (ASME, ASTM), and government agencies (NIST) have proposed standards for open data frameworks, systems, and ontologies that flexibly accommodate data for purposes of record keeping and analysis.³³ 

While pre-competitive knowledge sharing has been a norm in materials development, infrastructure for the sharing of raw data has only become broadly available over the last decade.³³ While such resources enable the rapid evaluation of materials at the discovery and selection stages, a standard has not yet been established for sharing manufacturing process data because data at later stages of the materials development process are often considered trade secret.³⁴ Over the past decade, science has progressed to a more open culture around raw data sharing, but that culture has not yet led to any widespread data sharing in the manufacturing sector. These challenges are substantial and will require cultural buy-in from multiple parties in order to realize the efficiencies at a global scale that would result from open data exchange. Of course, the motivation for such a sea change in industrial thinking must be derived from confidence in the technical feasibility of the entire enterprise, and there are nontrivial technical hurdles related to storing representations of materials data that are sufficiently flexible to be useful across the industry that must be addressed.

Most materials information, if structured at all, is stored in relational databases that impose a rigid schema on their data. For a problem of narrow scope, a relational structure can be feasible, but this architecture is fundamentally brittle: small changes to the original problem can precipitate large changes in the structure of the database or break the schema altogether and necessitate a new database for each use case. This is a problem of high visibility in the field of information technology, and multiple solutions exist under the heading of so-called “NoSQL” databases that store data in machine parsable documents with no externally imposed schema.³⁵ Materials scientists are already beginning to recognize the benefits of moving beyond a relational data model,³⁶ and we believe that this is only the first of many coming realizations that the Big Data revolution has ushered in technological advancements for which the materials industry has demonstrable need. For certain, the “materials genome” for describing metals, alloys, and polymers will all be dramatically different from one another, but conceptually there is no reason to prohibit comparing them on the basis of transferable properties such as yield strength, modulus, or toughness. In order to facilitate these comparisons, an open data platform with sufficient storage flexibility to hold all of the relevant data will need to be developed and embraced as a minimal first step, but the desired level of abstraction will be accomplished by embracing the extract, transform, load (ETL) model whereby a computation layer is deployed to translate disparate inputs into a common property format to facilitate comparison across material systems.

There are several government or nonprofit-led efforts to create testbed systems expressly for the purpose of enabling this degree of information exchange. These include “The Smart Manufacturing Project” run by the Smart Manufacturing Leadership Coalition and the recent Funding Opportunity Announcement for a Manufacturing Innovation Institute, both funded by the US Department of Energy.³⁷ In addition, Citrine Informatics offers private repositories for materials data spanning from development to manufacturing. These efforts, while in their early days, seek to create shared infrastructure, testbeds, and systems to motivate organizations of all types (e.g., universities, non-profit research labs, small businesses, and large corporations) to come together for common learning in precompetitive ways.

To date, materials informatics has in practice largely been focused on materials design and discovery on a laboratory scale. To truly leverage these discoveries going forward, the rest of the materials lifecycle must be considered holistically. This includes materials discovery, selection, and optimization for product use, certification, and manufacturing. Materials informatics—including data standards, infrastructure, and analytics—creates an opportunity to link all of these stages in compelling new ways. The coupling of theory, data, and experiment will enable faster development at all stages, but many specific questions remain unanswered. As the field continues to mature, spurred on by government, non-profit, and private efforts, opportunities for new analyses, method development, and rapid empirical data collection will only grow and lead to more efficient material deployment.

The authors would like to thank Professor Mark Johnson, Director of the Advanced Manufacturing Office of the US Department of Energy, for helpful conversations and perspectives in the preparation of this paper.