5

Emerging Applications

Haydn Wadley, University of Virginia, introduced three speakers who were invited to address emerging applications of data analytics in materials science: Apurva Mehta, Stanford University; Florencia Paredes, Citrine Informatics; and Carla P. Gomes, Cornell University. Ward Plummer, Louisiana State University, moderated a short Q&A following the presentations, and then introduced a panel discussion.

EMERGING APPLICATIONS: ACCELERATING DISCOVERY OF NEW TECHNOLOGICAL MATERIALS

Mehta discussed how artificial intelligence (AI) can speed the discovery cycle for new materials and underscored the important role of data in realizing this potential.

Speeding the Discovery Cycle

There are an untold number of elemental combinations that can be arranged in myriad three-dimensional (3D) structures (from short-range atomic arrangements to macroscopic structures) to create the new materials that will solve many of the world’s problems. Complex and hierarchical materials are crucial to meeting future materials needs. There are tens of millions or even billions of unexplored complex materials composed from just 30 common, nontoxic metals and metalloids in the periodic table. Yet with traditional experimental methods, finding and developing

even one or two of them into useful products would take over a thousand years, Mehta posited, and newer high-throughput methods would still require a decade. To bring promising new materials within reach on a much shorter time scale, additional guidance is needed. Mehta stressed that AI can provide that critical guidance to speed the discovery cycle. AI, he said, enables better predictions, smarter measurements, faster extraction of discoveries, and a means to reuse data (even from failures) to drive more discoveries.

Mehta argued that the current discovery cycle is broken, and AI is a vital tool for addressing this problem because the data are becoming too complex and come too fast to be manageable. As a result, valuable information is lost at each step, and experiments are done blindly, producing a lot of data but not much new information. He posited that AI, machine learning (ML), and other emerging techniques can fix this cycle, enhance the information gained from experimentation, and accelerate new materials discoveries.

These techniques have already improved predictions for materials design by linking disparate data sets. Mehta described an example in which ML was used to create a very accurate model for predicting metallic glass compositions, a complex material that requires at least 80 features (dimensions) for prediction. Current physics theories accounted for only a small fraction of the predictions in the model, and looking elsewhere for the necessary physics and chemistry knowledge, something machines can do better than humans, allowed the model to make better predictions.

Data Are Essential

Data are essential for ML/AI predictions to be accurate and useful, yet there are challenges to generating the right data, in the right locations (or data of sufficient quality and large diversity). The traditional model—testing one sample at a time and extracting data from it—is too slow. In the traditional search for new metallic glasses, for example, even when new materials were found, critical information capturing the diversity of the composition space (i.e., positive peaks as well negative valleys) was lost. Loss of the diversity, especially the negative valleys, made predictions less reliable. To address this, Mehta and his team turned to parallel synthesis and high-throughput characterization, allowing them to capture properties of a complete ternary of alloys in one day, resulting in fast learning and much more accurate predictions.

The final part of the discovery cycle requires extracting knowledge from the data. Raw data come in the form of images, but labels and classifications are essential to extracting knowledge from images and making new discoveries. Here, machines hold great potential, Mehta emphasized: Machines can be taught to extract information from images much faster than humans and learn on their own to extract information from noisy and highly multidimensional data sets that

overwhelm human perception. As experiments become faster and more complex, AI and ML will become ever more critical to help human researchers acquire new knowledge.

Mehta noted that for AI and robotics to function optimally, label data and knowledge must be stored for future use and made available to the whole materials science community quickly and easily. Mehta acknowledged that as the rate of data generation rises exponentially, it becomes increasingly more expensive to store data for future use. Mehta suggested that perhaps a good strategy may be to jettison some of the raw data from reproducible experiments as the data age, but carefully preserve conditions under which the data were collected (workflow) and associated metadata, so that much higher quality data can be recollected quickly when needed on improved instruments of the future. However, he argued that this must be done only with reproducible experimental data, and efforts must be made to preserve experimental data from rare objects and observational data from rare phenomenon.

Reiterating a key challenge discussed throughout the workshop, Mehta emphasized that knowledge is lost when data are not added to observational databases for recalculation and reuse. This critical step is hindered by the fact that negative results are not typically kept or published; positive results are often behind a paywall or in nonmachine-readable PDFs; intellectual property (IP) concerns limit sharing; and rigid formatting requirements can make data noninteroperable. For materials science to truly benefit from the AI and data revolution, materials science community must overcome these barriers and make data Findable, Accessible, Interoperable, and Reusable (FAIR), he said.

In closing, Mehta outlined two ways in which the materials community might organize itself in the years ahead: large groups of materials scientists with private reserves of data and proprietary analytics capabilities working in isolated siloes or a federated community of midsize groups sharing data and analytics.

Q&A

John Gardner, National Aeronautics and Space Administration (NASA), asked how researchers can know if the data on hand are the right data to use. Mehta replied that the right data will provide the density and shape of the information manifold plus the experimental dimensions. Predicting the manifold with sufficient accuracy will result in the desired knowledge. Another participant asked why databases for highly ordered phases were used to search for highly disordered phases. Mehta conceded that the information that exists is not always the information that is wanted, an important problem to overcome. He added that materials science data is relatively sparse, especially when considering its complexity; the sparsity is the major challenge to applying AI methods to materials problems.

Manoj Kolel-Veetil, U.S. Naval Research Laboratory, asked if boundaries between failure and success should be defined more precisely, and if there should be more incentives to share failures. Mehta agreed that negative results are much more important than the community concedes. Capturing these results must be prioritized because it comes with fewer proprietary concerns and will be an easy win, he argued. He added that boundary classification often requires domain (human) expertise, and experts do not always have consistent labels or thresholds for defining these boundaries. Uncertainty in definition of these boundaries contributes to variability and uncertainty in predictions from ML models.

June Lau, National Institute of Standards and Technology (NIST), mentioned that when AI has been applied to chess, human players often cannot figure out how the computer is winning. Maybe, she speculated, it is possible that AI can do something similar with physics, gaining new knowledge it can then share with us. Mehta agreed that this was possible, and noted that ML failures can also bring insights.

EMPOWERING SCIENTISTS THROUGH DATA AND AI

Paredes discussed Citrine’s philosophy and approach to addressing materials data problems and enabling materials discovery. Citrine is a materials informatics company that creates solutions for academia, government, and the materials and chemical industry. It is founded on a belief that AI and data technology can enable scientists and engineers to create new, useful materials. To accomplish this, Paredes emphasized that data, data tools, and data access must empower—not replace—scientists.

Case Study

Paredes described how an aircraft research center used Citrine’s platform to identify candidate materials for a light, durable, 3D-printable, aerospace-grade aluminum alloy. Their test alloy had unacceptable cracking, and researchers used Citrine’s platform to screen more than 11 million candidates and identify a short list of potential nanoparticles that could be added to eliminate the cracking. Within days, a viable alloy was found that could be used and commercialized.

This story illustrates how ML and other data tools can be powerful co-pilots in materials science, working with the domain experts who ask the right questions and generate the right hypotheses. Experts, with their essential domain knowledge, can use Citrine’s ML and data tools to perform years’ worth of work within a few days, Paredes said.

Materials Data Challenges

While data is essential for scientists and engineers to do their work, Paredes outlined several challenges to finding, generating, and using data for materials discovery.

First, materials data are messy. Scientists need to know the context of the data, such as how the data were acquired or what the specimen geometry was, before the data can be used, but such metadata is rarely available. Data are also often disconnected, even within the same company. Materials are frequently processed and measured by different groups using different formats, with reuse hindered by a lack of mechanisms to synthesize and unify the data.

To enhance data use and access, it is vital to consider what context is needed, when it is needed, and how the data should be structured and standardized. Citrine applies a data structure that captures the right materials metadata, connects important contextual information, and links the data in the right order. This structure is flexible enough to support multiple materials classes, yet rigid enough to enable the use of ML. In addition, data are recorded in a simple, open, and standardized format that facilitates data sharing and interpretability.

A second key challenge is that materials data are scarce. These data exist in small data sets, which could have a high sample bias; the data are sparse, lacking information about many measurements and properties; the data often do not span multiple materials; and the data require extrapolation, because they are rarely classified or labeled.

To solve these problems, Citrine applies domain knowledge to small data sets, enriching them with equations, simulation data, and experimental data. The company is also building capabilities to gather and leverage data and increase interpretability. Paredes stressed that Citrine wants scientists to be able to leverage more value from their data in order to better apply ML, better define their next experiments, and understand and trust the models.

A final challenge is that teams often lack effective collaboration tools to bring together the necessary materials and data experts for successful ML implementation. Scientific research is no longer a solitary pursuit; companies succeed when multiple experts share knowledge. The materials and chemicals industry needs to embrace this cultural change and encourage more collaboration, Paredes said. Citrine is building software and technology to empower collaborative teams, make useful and scalable tools that can be broadly applied yet solve specific problems, and encourage multidisciplinary collaboration and culture change.

Empowering teams starts with standardizing data to facilitate sharing and reuse. To that end, Citrine is creating a library of models, data sets, and other assets to make every project more efficient than the last. Such a library would enable materials scientists to apply their domain knowledge, build frameworks and

relational maps, leverage assets to identify properties, and build models to solve their specific problems, Paredes said.

Q&A

In response to a question from Robert Hull, Rensselaer Polytechnic Institute, Paredes clarified that Citrine uses all currently available AI tools to recover unstructured dark data from clients, although it can be challenging to place historical data in the right context. Another participant suggested Citrine could learn from LexisNexis, which is successful as a global aggregator of information because it pays for the data, and NIST, which has created a trove of vetted information on alloys.

WHAT THE FUTURE HOLDS: AI FOR ACCELERATING MATERIALS DISCOVERY

Gomes discussed AI’s transition from academic research to real-world applications, its potential for enabling discoveries, and examples of current efforts to capitalize on this potential, including CRYSTAL, a multiagent system for materials discovery.

AI in the Real World

After decades of research, AI has undergone a rapid shift from the academic realm to the real world over the past 10 years or so. With advancements in computer vision, speech recognition, and natural language processing, AI systems are starting to “see” and “hear,” and are even demonstrating superhuman capabilities in intellectual tasks like chess and Jeopardy. Semi-intelligent, semi-autonomous AI systems are being deployed in applications like social media, smartphone applications, and self-driving cars and are transforming business, medicine, and manufacturing.

The ability for AI to dramatically accelerate materials discovery will depend on fundamentally new, transformative approaches that incorporate physics and prior knowledge, Gomes said. Although materials discovery represents a highly complex, multidisciplinary computational problem, insights from other fields can help move the needle.

Today, materials scientists use ML to create materials databases and solve what Gomes described as relatively easy problems. AI’s next frontier, Gomes stressed, is to make new scientific discoveries. This will require an expanded toolset—for example, to incorporate reasoning by combining ML with “smart” or symbolic AI (representations of problems, logic, and searching). AlphaZero, the program that taught itself to play chess, did so by combining deep learning with reasoning. While current AI tools automate what humans can do without thinking, true scientific discovery requires thinking—something that is still out of reach for machines.

AI techniques have significant potential to yield new discoveries even in a field as complex as materials science, but there are many computationally difficult challenges to overcome. The first issue, raised throughout this workshop, is a general dearth of labeled data. In addition, Gomes emphasized the need to move AI beyond interpolation by predicting unknown materials and finding the right representations. Doing so will require incorporating prior knowledge and physics constraints, understanding the underlying phenomena, and managing uncertainty.

Demonstrating Value

Gomes discussed several current projects that use AI to enhance the work of human researchers.

One project, CRYSTAL, encapsulates many of the issues at the heart of the materials discovery challenge and AI opportunities in this space. CRYSTAL is a crystal structure phase mapping project that facilitates the use of high-throughput experiments for combinatorial materials discovery.¹ While synthesizing lots of materials is easy, inferring a material’s crystal structure based on X-ray diffraction patterns (XRDs) is difficult, creating an important bottleneck. This is because XRDs structural data can be quite complex and convoluted, blocking the ability to identify patterns and properties.

Examples of this “source separation” problem have been found in many contexts where a target must be extracted from a jumbled collection, including work focused on demixing overlapping sudokus, separating sources of music, and detecting bird calls. Standard ML techniques fail to capture the underlying physics and prior knowledge of these problems. However, by combining symbolic AI with ML and incorporating multiple data sources, prior knowledge, reasoning from density functional theory (DFT) approaches, and uncertainty, it is possible to produce a phase map that is physically meaningful.² For example, to discover new methanol oxidation electrocatalysts, CRYSTAL identified known structures within minutes and produced thousands of valid phase diagrams (although not all were configured correctly). Domain experts were then able to select the most promising structures for further study.

Another project, Scientific Autonomous Reasoning Agent (SARA), integrates materials theory, experiments, and computation. SARA is designed to encapsulate multiple steps of the scientific method for materials discovery, facilitating

___________________

¹ C.P. Gomes, J. Bai, Y. Xue, J. Björck, B. Rappazzo, S. Ament, R. Bernstein et al., 2019, CRYSTAL: a multi-agent AI system for automated mapping of materials’ crystal structures, MRS Communications 9(2):600-608, https://doi.org/10.1557/mrc.2019.50.

² S.E. Ament, H.S. Stein, D. Guevarra, L. Zhou, J.A. Haber, D.A. Boyd, M. Umehara, J.M. Gregoire, and C.P. Gomes, 2019, Multi-component background learning automates signal detection for spectroscopic data, NPJ Computational Materials 5:77, https://doi.org/10.1038/s41524-019-0213-0.

everything from database mining to hypothesis generation to experiment design. A key active learning component combines with data mining, probabilistic reasoning, and ML to help ensure that there is sufficient data.

In addition to CRYSTAL and SARA, Gomes is collaborating with the Toyota Research Institute (TRI) in a project that combines ML, symbolic AI, and human-computer interaction frameworks to facilitate catalyst discovery. She noted that many of the methods and frameworks for her materials discovery projects intersect with her other work in the field of computational sustainability, which uses computer science to address key societal and environmental problems.

Q&A

Kolel-Veetil asked Gomes to clarify how her projects incorporate the laws of physics. Gomes stressed the importance of including prior knowledge and constraints, while recognizing that they can be incomplete. She added that these problems are often so computationally intense that other challenges emerge, such as satisfiability. However, with the right approach and multiple iterations, structure can be found.

PANEL DISCUSSION

Plummer introduced the speakers for the workshop’s third panel discussion: John Gardner, NASA; Gareth Conduit, Intellegens; and James Warren, NIST. Haydn Wadley, University of Virginia, moderated an open discussion following panelists’ remarks.

John Gardner, NASA

Gardner stressed that new materials are useful only if their properties are predictable and manufacturable. He detailed a project, aimed at creating 3D-printable carbon nanotube yarn filaments, that demonstrates a path to building multifunctional materials and structures.

3D printing is very difficult to do correctly, and can easily create flaws in materials. Global parameter settings must be continually tweaked to eliminate specific errors, and then tweaked again to avoid subsequent problems. To overcome this variability, NASA created an end-to-end ML solution where the model can learn how to print a desired material correctly by optimizing the local printing parameter settings to minimize flaws, especially on the surface.

The model has three steps: error identification and classification (akin to quality control, and necessary because there are no data or literature to learn from); error prediction (parameter selection based on the gained error data); and parameter smoothing (which minimizes the possibility of failures). The resulting

materials were far better than those created with global parameter settings. In addition, the model created code—for example, for print speed, extrusion rate, and fan speed—that can be useful in future applications.

The process still faces challenges, however. In this case, the scope was quite general, and difficult questions remain for different geometries, including questions about what data should be collected, and how.

Gareth Conduit, Intellegens

Intellegens’s Alchemite ML platform tackles sparse data, a common problem in materials and drug design. As Conduit explained in his keynote address, standard ML algorithms exploit composition-property correlations, but Alchemite can predict from property-property correlations. It can do this without a strong material representation by creating maps to handle missing data and using first-principles computer simulations as input. Using Alchemite can dramatically reduce the number of experiments required, speeding up new materials discovery and achieving validation in as little as 2 years.

Alchemite has been used on several successful projects. Its data interpolation and extrapolation tools enabled the design and 3D printing of nickel-based alloys for jet engines and molybdenum-based alloys for forging. In addition, Alchemite was used to merge DFT calculations with experimental data to design welding consumables, create new metal-organic frameworks, and select additives for custom-made concrete.

Two different collaborations, focused on extracting data and merging sparse data sets, will further speed up the discovery process. With Biorelate LTD, Intellegens is creating a system that automates information extraction from images and scientific publications—in particular, for graphs and tables—to deliver high-quality, previously hidden data. Another effort, called Optibrium, has several aims: to extract all possible information about data, including insights on why data is missing and what is hidden in the noise; to develop ML models that better extrapolate data capabilities; and to merge models that have been trained on private data silos.

James Warren, NIST

Warren is the director of the Materials Genome Initiative (MGI), an effort to accelerate the discovery, design, and deployment of new materials. MGI has laid the groundwork for a materials innovation infrastructure that can achieve national goals in the face of global competitiveness and educate the next generation of materials workers.

AI and physics are alike in that they are both ways of generating models that are useful, but also provisional and uncertain. Experiments that can test

models contribute to ground truthing and also generate new knowledge, especially in sparse systems. For example, integrated computational materials engineering (ICME) models use databases and process-structure-property (PSP) linkages, and while AI is not essential for ICME, it can be inserted into any step, even without a physics-based model, to speed up the process.

For new knowledge to be generated, Warren emphasized that data sharing needs to be simplified and encouraged via a sea change in data sharing tools and culture. In addition, today’s “state-of-the-art” tools—high-throughput computation and experiments in semi-autonomous laboratories—must evolve into a policy-coordinated, integrated data infrastructure with more autonomy, more direct incorporation of physics, and more interpretable AI.

OPEN DISCUSSION

Participants and panelists had a lively discussion on needs for advancing AI and on the tension between trusting AI and understanding it.

Needs for AI in Materials Design

Participants discussed multiple elements that are needed to advance the use of AI in materials design, including data, computation, time, and instrumentation.

Tresa Pollock, University of California, Santa Barbara, asked whether it is necessary to mine legacy data. Gardner said that legacy data will remain important, at least until autonomous experiments improve and make it more feasible to simply generate new data. Beyond that, the answer depends on the application. Warren added that there will always be a role for data, but we should also be automating as much as possible to focus on better questions, and it will likely be a mix of robotics and brute-force experimentation that moves things forward.

Mehta stated that AI needs both data and computation, but as we move toward computation-informed experimentation or manufacturing in real time, low latency and the ability to handle higher-intensity computation will become increasingly important. Warren agreed, and Gardner underscored the importance of latency time in manufacturing. While preemptive data processing can help, today’s machines are not always capturing the kind of data necessary to support real-time feedback.

Gomes pointed out that active learning is important to reduce search by several orders of magnitude, but said that it is unlikely that in the near future we will be able to just use active learning to solve scientific discovery problems quickly: We need to fundamentally advance our general algorithms. For example, the current state of the art for crystal-structure phase mapping (including active learning) is far from solving the problem for arbitrary instances, even if we have

lots of data and a lot of time. The process can solve only very simple instances of crystal-phase mapping. Getting to the desired level of efficiency will take time and require AI research combining reasoning (about prior knowledge) and learning algorithms, she said. Conduit added that while it takes a long time to train a model, once it is ready it can make predictions very quickly. He suggested that it may be possible to reduce an ML model to its core to simplify predictions and improve data interpretability.

In response to a participant, Gardner noted that 3D printers can be a limiting factor now, but machines are constantly improving their data collection, although new features must be balanced with manufacturing capabilities and budgets. Warren added that the larger issue is that all instruments, whether 3D printers or electron microscopes, have to work with the ML and physical models and handle uncertainties, respectively. Conduit agreed and noted that it is important that machines and models be able to tolerate acceptable ranges. Gardner added that localization is also critical to ensure that data matches up with reality.

Wadley noted that Bayesian methods could infer properties and statistical distributions, and wondered if they could be used to optimize ML for materials design. Gardner agreed, and Surya Kalidindi, Georgia Institute of Technology, added that design factors are often used to combat ignorance and increase safety. Reducing the number of design factors, he continued, will work only with more information about the material and new techniques to assess its structure.

Trust Versus Understanding

Participants discussed the merits of understanding why an AI model works versus simply trusting the results.

Broadly speaking, Gardner posited that building trust in AI is a challenge that will take time, but successful results speak for themselves. Warren agreed that AI faces perception problems, but noted that all models can have errors—even the physics models researchers rely on are wrong to some extent. Sometimes the proof comes from validated results, and other times it is a matter of determining one’s level of comfort with how the results are derived at each stage. Building on this, Conduit underscored the importance of capturing the uncertainty in the predictions of any model.

Hull noted that trust in AI will depend to some extent on the interpretability of results. Kalidindi expressed his view that interpretable and trustworthy are very different, and suggested that scientists should trust nothing and work to validate everything. Gomes added that improving interpretability creates new advances, because humans and models will need to work in concert. Making progress in this area will require better mechanisms for extrapolation, integrating physics, and incorporating prior knowledge, but this effort will be worthwhile, she said,

because improving AI in the context of materials science can also contribute to major advances in AI for other applications.

If the results are useful, Mehta wondered, when is it necessary for scientists to understand the models? Gardner noted that from a manufacturing standpoint, all that matters is that the output is verified. Conduit agreed that understanding the model is not essential if the material ultimately performs well—indeed, in the future, the models may be so effective that they are considered their own verification. Gomes agreed, but stated that it remains essential to conduct a separate verification of the solution to create trust in the model.

On the other hand, Warren said that a better understanding of how AI models work could expand materials science knowledge in general. Right now, correlation is the best available tool for understanding how materials work. Conduit stressed that there are many situations where an understanding of the underlying principles is needed, and that developing this understanding can in turn inspire creative solutions.

Lourdes Salamanca-Riba, University of Maryland, noted that the tension between understanding and trusting AI is at its core a cultural issue, and one that may be perceived differently by the emerging cohort of researchers as compared to the previous generation. Gardner noted the relationship between this issue and the broader cultural evolution in terms of how people decide what information is “real.”

Another participant asked what was more important: new tools to solve problems, or advancing science overall? Warren replied that the answer depends on who is asking, and if there is a deadline. Gardner agreed, noting that those focused on enabling manufacturing would prioritize problem solving, whereas experimentalists and scientists are more likely to pursue overall understanding for better systems design.