Autonomous Materials Discovery and Optimization
Proceedings of a Workshop—in Brief
On November 1–2, 2022, the Defense Materials, Manufacturing, and its Infrastructure (DMMI) Standing Committee of the National Academies of Sciences, Engineering, and Medicine held a workshop on autonomous materials discovery and optimization. This Proceedings of a Workshop—in Brief summarizes the presentations and discussions that took place during that workshop.
INTRODUCTION AND OVERVIEW
DMMI chair Haydn Wadley from the University of Virginia noted that the development of new materials has historically been crucial to growing the economy and has enabled new capabilities that improve the health of our citizens and the security of our nation. However, the time and cost to develop materials has grown and it is not unusual for it to take millions of dollars and decades of work to bring a single new material to market.
Workshop planning committee chair Andrea Hodge of the University of Southern California added that this meeting would bring together diverse experts to discuss how to integrate autonomous materials discovery and optimization into the research and development process. In contrast to automated systems, which perform predetermined tasks as specified, autonomous systems can learn, adapt, and change their performance without human input.
Adam Rawlett of the Army Research Laboratory predicted that autonomous materials discovery will accelerate the pace of materials science and fundamentally change the way materials science is done across sectors. Speeding up the search for and development of materials with specific properties will have important implications for the entire world.
KEYNOTE
Alan Aspuru-Guzik of the University of Toronto described autonomous materials discovery and optimization in molecular biology and the lessons from that field that can be applied in materials science. It generally takes 20 to 30 years for a biological material to go from discovery to commercialization and scale up. The key to speeding up this discovery and development process is what Aspuru-Guzik called the “self-driving lab”—a machine that analyzes data from samples that have already been made and uses those data to decide what material to make next. He spoke of two major challenges related to developing self-driving laboratories.
First is the designer challenge—moving from specifying the desired properties of a material to defining the structure of a material that will result in those properties. He said self-driving laboratories need to take advantage of that developing capability of artificial intelligence (AI)-driven generative models for materials design. It will be important to think about the constraints that are being programmed into these models so that one can optimize for desired material properties and other considerations such as their environmental impact, cost-effectiveness, scalability, and ethical considerations. They should also be iterated with real data to ensure accuracy.
The second major challenge discussed was the development of the “matter computer”—that is, a technology that can create a desired material from basic building blocks. He described work that his laboratory has done in finding an organic molecule similar to an organic light-emitting diode (OLED) but that lases so that crystals of the molecule could act as an organic laser. Because such a laser would operate with lower power requirements than solid-state lasers, it could have many possible uses. Taking advantage of AI search techniques, his group identified a number of compounds with much better properties than the original lasing OLED.
Ultimately, he concluded, self-driving laboratories should be able to carry out the discovery process, create devices with the materials that have been synthesized, and test those devices.
AUTONOMOUS MATERIALS DESIGN
The day’s first session was led by workshop planning committee members Susan Sinnott of The Pennsylvania State University and Klavs Jensen of the Massachusetts Institute of Technology (MIT) and consisted of two main presentations plus discussion by a three-person panel.
Additive Manufacturing
Ryan Dehoff of Oak Ridge National Laboratory (ORNL) spoke about work on additive manufacturing (AM) techniques that will make it possible to create complex metal structures, such as turbine blades. The focus has been on discovering which of the many possible processing approaches will lead to the best result—and that will not crack or otherwise fail under intense operating conditions. This work is very hands-on and requires a great deal of input from engineers and materials specialists, but autonomous approaches could play a major role in the future.
AM is a process by which an item is created layer by layer in the desired shape, in contrast to subtractive manufacturing, where an item is carved and routed out from a larger piece, either a solid block of material or a piece that has been forged in a shape that approximates the finished product. One AM technique involves laying down a thin layer of powder in the desired shape and then directing a laser beam at the powder to melt it and produce a layer of solid metal; another powder layer is placed on that, he explained, with the laser melting it and also welding it to the preceding layer. The process is repeated many times—sometimes thousands of times—until the item is complete. There are many different parameters that can be adjusted in the process, such as the speed of the laser moving across the powder or the pattern that the laser follows to cover the entire shape, and the choices made for these parameters will affect the grain structure of the finished metal product, which in turn will determine the item’s mechanical performance.
AM can provide incredible control over an item’s mechanical properties, Dehoff said. For instance, turbine blades created by the standard casting process will fail in unpredictable ways, depending on the random details of the cast item’s production. However, Dehoff’s team found that blades produced by AM failed in consistent ways. This ability to control the blades’ properties made it possible to optimize the AM process to produce blades that were extremely resistant to cracking.
Dehoff concluded that optimizing AM processing conditions for a particular design involves producing the same item over and over again in hundreds or thousands of different ways, examining the microstructures of the different items and testing their performance, and then combining those data with a theoretical understanding of the production process to create a model that can predict the microstructure and performance that will result from any given set of conditions. Although the analyses
that the ORNL researchers have been performing require a great deal of human input, ultimately it should be possible to do much or all of the work autonomously.
Developing Photovoltaic Technologies
Christoph Brabec of Friedrich Alexander University in Erlangen, Germany, described efforts there to develop autonomous methods for designing efficient photovoltaic devices, including the optimization of the design and production of photovoltaic materials.
To meet clean energy goals, he stated that it will be necessary to improve photovoltaic technologies at a much greater learning rate than is typical with conventional methodologies. So, how can this learning rate be accelerated? A major part of the approach that Brabec’s team is taking to accelerate the development process is to automate the manufacture and testing of photovoltaic devices. Their AMANDA Line 1, for example, is an automated device line that can start with basic materials, build a device, and then test it to create all of the data necessary for understanding the performance of the device over time. The line, which includes about 150 robots that communicate with one another, is set up like a classical device processing line, with solution processing, thin film deposition, metallization, quality control, characterization of the solar cells, and aging. Starting in the morning, the line will take about 8 hours to process 250–300 devices, and accelerated lifetime testing is carried out overnight. Each of the devices is built differently from the others, so within 24 hours it is possible to characterize the performance of 250–300 variations of a device. The resulting data are used to understand how variations in design affect performance.
In a separate line of research, the team has used the automated line to optimize a perovskite semiconductor for use in solar cells and other applications. The best perovskite semiconductors known today contain lead, so the team is looking for lead-free materials that will have optimal photoelectric performance, as well as perovskite compounds with other characteristics. The properties of the ultimate device will depend on its design and fabrication. Optimizing the device will require making and testing many variations, and Brabec’s team has built an automated device to optimize the processing of perovskite thin-films devices.
He concluded by noting the team has a Bayesian optimizer that uses AI techniques to make suggestions for which materials or devices have the most promise and should be tested with the automated lines. The goal is to create a fully autonomous system in which the results of autonomous testing are fed into the AI system, and the system decides which experiments to do next according to predictions based on the data accumulated from all of the previous runs.
Panel Discussion
Artificial Intelligence/Machine Learning–Based Materials Design at Corning
Adama Tandia of Corning Incorporated described how Corning has used AI and machine learning (ML) to speed up its materials design. Their typical design process involves finding a material that meets specifications provided by a customer, including the material properties, raw materials, and manufacturing constraints. Designers then use multi-objective optimization to find a material that best meets the collection of specifications. Tandia explained that the company found that a physics-based ML model that combined neural networks with physics knowledge could produce extremely accurate temperature–viscosity graphs, which in turn could be used in materials optimization.
Historically, Corning has taken many years and millions of dollars to design a single glass composition, but the company came up with its new glass, Astra, in 6 months and at a fraction of the normal cost. This was done by using ML to build accurate models for each of the desired properties and putting them all together into a multi-objective optimizer, which produced a list of 40 reasonable candidate materials.
In closing, Tandia noted the goal is to embed the developmental aspects of creating a new material into the research phase, thus considerably shortening the process of creating a new material. The company may now be able to deliver a new product in as little as 3 to 4 months, he said.
Autonomous Design and Machine Learning for Materials Applications at the National Institute of Standards and Technology
Francesca Tavazza of the National Institute of Standards and Technology (NIST) provided an overview of efforts at NIST to use autonomous design and ML in the design and characterization of materials. AI has been a major focus at NIST for several years, she said, with dozens of AI-focused projects; Tavazza focused on efforts to apply AI research to materials measurement.
Tavazza touched on several specific AI-related materials projects at NIST. The first was the Closed-Loop Autonomous System for Materials Exploration and Optimization (CAMEO), which was described in more detail by Gilad Kusne on the workshop’s second day. One application of CAMEO was to find a best-in-class phase-change memory material, which was done 10 times as fast as with standard methods. A second example was the Autonomous Neutron Diffraction Explorer (ANDiE), which uses neutron diffraction measurement coupled with physics constraints and a Bayesian inference loop to study magnetic transitions and identify where the transition occurs and what the mechanism is. ANDiE speeds up materials search by a factor of 5.
Tavazza also mentioned JARVIS-ML, the ML component of the Joint Automated Repository for Various Integrated Simulations (JARVIS) project, which has collected a suite of databases and tools. JARVIS-ML has been put to work on a variety of materials-related projects. Dissemination is important at NIST, she said, so the institute makes everything it does available online, including CAMEO and JARVIS-ML. Outside researchers can gain access to databases, codes, workflows, and AI tools.
Autonomous Materials Discovery at the National Science Foundation
Cosima Boswell-Koller of the National Science Foundation’s (NSF’s) Division of Materials Research spoke about various NSF programs focused on autonomous materials discovery.
About a decade ago, NSF developed the Materials Genome Initiative for Global Competitiveness Strategic Plan, MGI 1.0, which was intended to bring together computational methods, experimental work, and digital data with an iterative feedback loop. MGI 2.0 was released in November 2021 with a focus on significantly shortening the time required to discover a material and bring it to market, potentially using AI, ML, and autonomous materials discovery.
Among the NSF programs that support research into autonomous materials discovery, Boswell-Koller said, the materials research science and engineering centers are some of the most important. Recent solicitations for these centers encouraged integrating theory and experimentation, and one of the five specific focus areas listed in the solicitation was AI, including the use of ML, deep learning, computer vision, and other emerging data-centric approaches. A related program supports materials innovation platforms.
Boswell-Koller concluded that the Materials Laboratories of the Future program was developed to provide direction on what future materials laboratories should look like and to accelerate the unification of the materials innovation infrastructure. The program envisions AI and ML as playing a major role in those laboratories of the future.
TESTABLE HYPOTHESIS APPROACHES
The day’s second session was led by workshop planning committee members Kelly Nygren of Cornell University and Edwin Thomas of Texas A&M University–College Station and consisted of two main presentations plus discussion by a three-person panel.
Translating Big Data from Simulations into Experimental Hits
Rafael Gomez-Bombarelli of MIT provided an overview of how using the results of first-principles calculations as data for ML can offer a powerful approach to learning about materials. He emphasized that each technique has its advantages and disadvantages. First-principle calculations, for example, require little or no fitting, allow extrapolation, and are high-throughput, but they are only as good as the model. ML is fast, takes advantage of large data sets, and can have uncanny performance, but it is only as good as the training data, which can be limited and expensive.
He offered a couple specific examples of this from his research group. In high-throughput virtual screening, there are often too many candidate compounds to
screen. To address this, a fraction of the compounds are simulated, and the remainder are screened with an ML surrogate of the simulator. Once a handful of promising candidates are selected, they are presented to subject-matter experts who help with the final selection.
A second example involved zeolites, a class of nanoporous materials—that is, materials containing nanometer-scale holes and tunnels—that are widely used as catalysts. Zeolites are very appealing for ML applications, Gomez-Bombarelli said, because there are almost countless variations whose structure depends intimately on their composition and processing. This provides a broad design space through which algorithms can search to find materials with particular properties, such as being well suited to catalyze a particular reaction. Nearly 3 million zeolite structures have been predicted, 252 have actually been made, and about 10 of those are used commercially. The MIT group ran a huge screen over all of the known possible zeolites, seeing which ones were best for particular uses.
Representation learning is key to this sort of ML-based analysis, Gomez-Bombarelli said—in particular, how certain data structures are input into an ML model. In the case of images, for instance, convolutional neural networks work well. Atomistic systems, on the other hand, require graph-based neural networks. He said that a particularly valuable type of model is the multi-fidelity model, which takes advantage of the strengths of both theory-based computations and experimental data. Multi-fidelity neural networks typically produce more accurate results than either non-AI models or the traditional neural networks.
Dealing with Bias in Autonomous Learning Systems
Jason Hattrick-Simpers of the University of Toronto offered some caveats on AI and autonomous learning based on how human biases can influence AI-based analyses. The community has blind spots when it comes to data interpretation, he said, which affect the data that go into the databases used to build models. As a result, AI is less a lens into the future than a mirror that “integrates our cumulative biases and errors and then reflects them back at us.” Given that, his talk provided a broad perspective on where human bias, uncertainty, and error can sneak into the autonomous workflow, particularly through the data used.
Hattrick-Simpers said that human analysis of spectral data is highly subjective and prone to optimistic interpretation. As an example, he spoke about the interpretation of electrochemical impedance spectroscopy (EIS). The typical way to interpret EIS is to build equivalent circuit models and find one that accurately reproduces the data. Many different models may provide reasonable fits, and the judgment of which provides the best fit is prone to experiential bias and strongly relies on tacit knowledge. One way to deal with these issues is to use AI and ML to create a large number of potential equivalent circuit models and then consult with experts about which make the most sense instead of simply looking for the best numerical fit.
Bias in data sets can arise in various ways, he noted. For instance, experimental data sets are not randomly chosen from the entire research space; they come from experiments that researchers chose to do and whose results were chosen for publication or inclusion in a data set. It is difficult to specify exactly what the selection bias in such a process might be, but it is likely there. Similarly, he said that data in computational data sets generated from AI and ML can be biased by the choice of models used to generate the data. One way to deal with this is to look at data from multiple models, as disagreement between ML models can uncover previously unknown biases.
Hattrick-Simpers concluded that the use of AI in autonomous learning in creating and testing hypotheses is very much a human endeavor despite the emphasis on machines. Thus, in developing AI models for data analysis, he suggested that people specify constraints, utilize statistical tools when possible, and create human–AI interfaces that provide users multiple views of potential models.
Panel Discussion
The Future of Autonomous Experimentation
Benji Maruyama of the Air Force Research Laboratory (AFRL) discussed the future of autonomous
experimentation. He emphasized the need to move more quickly in developing materials; the decades-long delay between the discovery of a material and its commercialization at scale impedes the impact that these materials could have. He also noted the need to attract more U.S. students to the workforce to support autonomous materials discovery and development; diversity, equity, and inclusion are critical.
Maruyama mentioned AFRL’s development of the Autonomous Research System, the first autonomous closed-loop research robot for materials, which learned how to grow carbon nanotubes on its own. Autonomous research robots are revolutionizing how research is done by greatly expanding the amount of work that an individual scientist can do.
Ultimately, he concluded, the best approach to develop autonomous experimentation may be with public–private partnerships among government, industry, and academia.
How Automated Material Synthesis and Testing Can Help the Defense Industry
Andy Detor of the Defense Advanced Research Projects Agency described several challenges facing the defense industry that could be addressed by automated and autonomous material synthesis and testing.
The first issue Detor raised was the need for adaptive hardware. This can be a major challenge to a product when its environment changes, either the geographical environment in which it operates or the logistical environment in which it is developed and built. As examples, he mentioned the problems with sand erosion of rotor blades encountered by the Black Hawk helicopter when it began operating in a desert environment and the materials design challenges posed to the F-35 Joint Strike Fighter by the wildly changing cost of rare elements such as rhenium, because each plane requires nearly 1,000 pounds of those elements. To deal with these challenges, Detor said, it is important to design adaptive hardware that can be quickly redesigned to meet those changing conditions.
That in turn requires engineers and materials scientists to work together on design, but that is challenging because it generally takes a decade to design a new material, so materials scientists end up guessing what engineers might require a decade later. Another materials-related design problem is that because most components are made from a single material, the material must be chosen to stand up to the most extreme conditions any part of that component experiences; the solution could be to use different materials for different sections of a component to create a body that is designed to optimize both material and shape.
This is where autonomous material synthesis and testing could make a big difference, he said, by rapidly generating data sets of design-relevant properties. If that can be done, then ML or other surrogate models could use those data to determine jointly the best material and the best design for a given objective. The key would be to have enough of the right sort of data to be able to optimize on both the material and the design at the same time.
Discovery Versus Optimization
Horst Hahn of the University of Oklahoma and the Karlsruhe Institute of Technology offered some comments on how best to use autonomous materials synthesis and testing.
His main concern centered on the role that intuition has traditionally played in materials science and what role it might play in a future where autonomous machines are responsible for much of the exploration of new materials. He noted that scientists’ intuition have led to the discovery of radically new materials, such as carbon nanotubes, nanocrystalline materials, bulk metallic glasses, and high-entropy materials. It is unlikely that a computer or algorithm would have found any of these, he said.
Hahn noted that these materials point to a way in which autonomous synthesis and testing can play a very important role, and that is in the optimization of such materials once they have been discovered. There is, for instance, a vast space of high-entropy materials (i.e., those comprising five or more elements), and this is a space that ML and AI are ideal for exploring.
SECOND-DAY KEYNOTE
Jed Pitera of IBM’s Almaden Research Center spoke about efforts at IBM to discover valuable new materials more efficiently and quickly and at a lower cost. IBM’s focus is mainly on materials used in semiconductor manufacturing, although the approaches being used there could be applied to materials in any field. Currently, it takes about 10 years and $10 million to $100 million to develop a material for use in semiconductors; IBM would like to cut down both the time and the cost of that discovery by 90 percent.
A number of powerful technologies can be applied to accelerate this discovery process, Pitera said, including supercomputers, AI, and the forthcoming quantum computing. Pitera mentioned four classes of tools that can be used in discovery. The first type are deep search tools that comb through the scientific literature to create a database. The second are AI-enriched simulation tools that screen and augment databases with predictions of material properties and chemical reactions; these simulation tools are becoming much more autonomous, he said. Third are hypothesis-generating tools that can generate candidate materials and molecular design. Last, laboratory automation and autonomous tools can generate materials quickly and consistently and test their properties. It is important to surround these tools with open infrastructures to build communities of researchers that not only use the tools for their own purposes but also add on to them so that the tools are constantly being enriched and improved.
Pitera described work at IBM to find a substitute for a photoacid generator (PAG) that is used in semiconductor manufacturing but is being phased out because of its toxic properties. Looking for a safe material with similar performance, the IBM team searched through some 5,000 known PAGs for those with suitable absorption properties, toxicity, and biodegradability. Because not all of the chemicals they found in their search had data on all of the important properties, they filled in gaps in the composition space with computations. They brought in human experts to filter the candidates that had been identified in the search and also used a surrogate model for human experts because there were too many candidates for the experts to sift through alone. After zeroing in on a list of suitable candidates, they made samples of the materials with a robotic line in order to test them. The entire process was finished in 9 months, versus what Pitera said would have been 2 to 3 years using the normal processes.
After describing two other examples of ways in which the various technologies are being put to work in the search for new materials and mentioning a number of emerging technologies such as integrated risk assessment, multi-fidelity simulation screening, and automated polymer synthesis and characterization, Pitera closed by describing key gaps and challenges in the field.
The first he mentioned is how best to define and represent a problem. There are good models for optimization, he said, but not so much for exploration. It is particularly difficult to encode “crazy ideas” that might work, such as new mechanistic or materials performance hypotheses. A second challenge is how best to manage the trade-offs between engineering and discovery. He noted that IBM has many legacy tools that were built for a particular capability and may be one of a kind in the world. It can be very expensive to automate these, so it is not clear if they should be included in the broad effort to implement automation in discovery work.
Pitera concluded that there is an issue of explainability and acceptance of autonomous discovery efforts. When these efforts bear fruit, humans must get involved in looking at the output of the automated or autonomous system and making some decision. Thus, it is necessary to be able to explain and also constrain the choices that the system has made with input from knowledge and subject-matter experts.
AUTONOMOUS CHARACTERIZATION AND PROPERTY MEASUREMENT TOOLS
The day’s first session was led by workshop planning committee members Robert Hull of Rensselaer Polytechnic Institute and David Aspnes of North Carolina State University, and consisted of two main presentations plus discussion by a four-person panel.
A Powerful Tool for Characterization and Property Measurement
Following Pitera’s keynote, Anna Lena Eberle of Carl Zeiss MultiSEM GmbH described the MultiSEM, a powerful scanning electron microscope (SEM) tool developed by Zeiss to study materials of various sorts. Instead of the single electron beam in a standard SEM, the MultiSEM has many parallel electron beams that carry out simultaneous scans to create a complete image of an entire area. The multiple beams make it possible to get an atomic-scale image of a much larger area than can be done with a single-beam SEM.
The tool is automated, Eberle said, so it can run unattended for hours or days. As an example of its power, she showed an image of a 1 cm by 1 cm semiconductor test chip. A state-of-the-art traditional SEM would take 4 weeks to create an image of the entire chip, while the MultiSEM was able to do it in 7.
MultiSEM has also imaged different battery components from different points in the battery’s life cycle to observe how the materials in the battery changed with use. An advantage of the MultiSEM, Eberle said, is that it provides a high-resolution image of a large area, so researchers can identify points of interest and zoom in to see details at nanometer scale. Ultimately, she said, the goal is to develop automatic image analysis algorithms to investigate such samples without human direction. In other work, researchers used MultiSEM to create large, high-resolution images of a metallic sample before and after mechanical deformation and to detect and characterize the hundreds of thousands of individual slip events created by the deformation.
Eberle emphasized that MultiSEM makes it possible to quickly create nanometer-scale images of large areas of a material without human monitoring.
Machine Learning in the Study of Ferroelectric Materials
Nazanin Bassiri-Gharb of the Georgia Institute of Technology and NSF provided a detailed example of how ML is being used in the study of ferroelectric materials. The field has not yet reached the point where autonomous learning is possible, she said, but the sort of ML she discussed could at some point be made autonomous.
Ferroelectric materials, Bassiri-Gharb said, are crucial in the manufacture of underwater sonar equipment and medical ultrasound transducers. They are piezoelectrically active, which means that applying strain to them results in the production of an electric charge across the material and, vice versa, applying an electric charge causes them to deform. They also spontaneously polarize electrically, and the direction of the polarization can be switched by applying a strong enough electric field. Most importantly, this polarization switching is nonlinear and follows a hysteresis curve, so the forward and backward switching proceed along different paths and are not simply the reverse of each other.
It can be difficult to examine the properties of these materials, she said, because many unconventional ferroelectric materials cannot be probed with conventional means. Instead, the probe of choice is a piezoresponse force microscope, a variant of the atomic force microscope that gets data on domains in a ferroelectric material by applying an alternating current (AC) electric voltage to the surface of the material with the tip of the microscope and measuring the resulting local deformation. A modification of that approach where the AC voltage is applied at a range of frequencies (called band excitation piezoresponse force microscopy) provides a great deal more information about the material under study. These measurements result in a large amount of complex data involving tensor values, and Bassiri-Gharb’s group has been using ML to correlate the data with material properties.
In closing, Bassiri-Gharb stated that her field is far behind much of organic chemistry, where much of the work, from synthesis to analysis, is done autonomously. This area cannot synthesize samples fast enough to create large data sets, and the tools are not available to do all of the measurements they would like. Data curation is crucial, particularly the encoding of any known correlations within data sets, but that requires the participation of scientists who are familiar with the field and can curate the sets properly.
Panel Discussion
High-Throughput Rapid Experimental Alloy Development
Kenneth Vecchio from the University of California, San Diego, described a system that he developed to rapidly characterize bulk samples of alloys. The system, named High-Throughput Rapid Experimental Alloy Development, includes a synthesis component and multiple testing components.
Vecchio said that the synthesis component uses directed energy deposition to melt powders into bulk materials of the desired shape and composition. The individual samples are made on a disk and are generally in the shape of individual slices of a pie or thick spokes running from the center of the disk to its outer edge. This circular symmetry makes it easy to automate both the synthesis and the testing, as simply rotating the disk allows a machine to move from one sample to the next.
Vecchio has a variety of different machines that characterize the samples using X-ray diffraction, SEM, electron backscatter diffraction, nanoindentation, glow discharge spectroscopy, and other techniques. The synthesis and the measurement techniques are individually automated, but they each need to be set up by someone placing the sample disk in the machine.
Perspectives on Autonomous Materials Discovery
Brad Boyce of Sandia National Laboratories (SNL) discussed the challenges inherent in autonomous materials discovery. A major issue driving the push for autonomous discovery is the “curse of dimensionality”—that is, the vast number of factors that must be taken into account in any attempt to optimize materials and processes. Boyce showed a slide listing more than two dozen process parameters, another two dozen factors related to a material’s structure and chemistry, and yet another two dozen properties and performance factors that may be important for a given material. All of these may need to be taken into account when optimizing a material, he said, so, given the nearly endless number of combinations that are possible when looking at these factors, how will it ever be possible to truly be certain that one has identified optimal solutions? Making matters worse, he added, materials are stochastic, so each time a sample is produced, the thermodynamics and kinetics act in slightly different ways. This led him to high-throughput testing, and, like Vecchio, he has automated about a dozen instruments for materials characterization, including X-ray diffraction, X-ray fluorescence, SEM profilometry, and nanoindentation. Furthermore, he has developed a robotic arm that can move samples from one testing device to another.
He has also become interested in what he calls proxies, or “surrogate low-cost fingerprints for materials.” As an example, he described how a postdoctoral student used photographic images of 25 three-dimensional, printed stainless-steel lattices that looked very similar but were printed under different process conditions and absorbed very different amounts of energy under compression. When these photographic proxies were analyzed by an AI algorithm (a neural network), the algorithm was able not only to successfully predict the compressive deformation properties of the lattices but also to provide insights that allowed Boyce’s team to develop a hypothesis explaining why the different lattices performed differently.
Last, Boyce said that a group he is leading at SNL has a 3-year, $15 million project called Beyond Fingerprinting that is aimed at exploring material process optimization in a multi-modal data integration space. This is the sort of tool, he said, that could prove very powerful in exploring materials.
Artificial Intelligence in Materials Testing
Brian Sheldon of Brown University offered two examples of using AI in materials testing and then provided lessons from his work in materials synthesis and testing.
First, he described measuring the mechanical behavior, such as fracture toughness, of ceramics and nanocomposites by applying pressure to a micro-cantilever made from the material of interest. One can calculate an approximate solution of the fracture toughness using simple beam bending theory, but this is often not accurate enough. On the other hand, finite element model (FEM) calculations are accurate, but they are not practical for a large number of measurements because they take too long. Sheldon’s group got around this problem by training an ML model on the FEM data; the model then produced values for new tests that were
as accurate as the FEM calculations but took much less time to produce. His group did something similar with pillar splitting experiments used to measure fracture toughness in ceramics.
Sheldon ended with two lessons he learned from a 20-year collaboration between Brown and General Motors on computational materials science for developing materials for use in the automotive industry. First, computations can be done more quickly than the experiments—automation and high throughput can help bridge this gap. Second, electrochemically materials such as those used in batteries are messy to work with, and could be challenging to develop such materials autonomously.
Building Trust in Novel Materials Created by High-Throughput Research Tools
Tonio Buonassisi of MIT began by offering a concrete example of materials discovered by autonomous research. Dipping two polymers into hot water, he showed how one became cloudy and the other remained clear. They had been designed via ML to have different cloud point transition temperatures and had then been developed in a high-throughput cycle. The goal, he said, is to have many more such success stories in the future.
To illustrate the power of high-throughput research tools, Buonassisi described two discoveries. The first was the development of a single compound, methylammonium bismuth iodide, which took 6 months to create and whose structure and properties were characterized in great detail. The second involved the high-throughput synthesis and characterization of nearly 100 compounds over just 2 months; those included two new materials that had been unreported in the literature and also four that were grown in thin-film form for the first time.
However, there was much less characterization done of the materials’ structure and properties, leaving what Buonassisi described as a “characterization deficit.” This is a problem with many high-throughput loops and needs to be addressed. One of the ways to overcome these deficits would to be to use proxy-based measurements, as mentioned by other speakers.
Looking to the future, Buonassisi described an effort at MIT to use high-throughput inkjet printers to create thousands of unique compositions per minute. However, such a new technique raises questions about how much the results can be trusted, and he said that his team is trying to build confidence by comparing the material properties of the compositions created with the inkjet method to those of materials created by the traditional approaches.
SMART PROCESS TOOLS
The day’s second session was led by workshop planning committee members John Koszewnik of Achates Power, Inc., and Lourdes Salamanca-Riba of the University of Maryland, and consisted of two main presentations plus discussion by a three-person panel.
Robotics for Use in Autonomous Materials Synthesis and Manufacturing
Satyandra K. Gupta of the University of Southern California provided a review of the current state of robotics and discussed the use of robots in autonomous materials synthesis and manufacturing. He broke the capabilities of current robots into a number of different categories.
First, he said, advances in learning are enabling robots to program themselves from high-level task descriptions. Such advances include robots now being able to learn by watching a demonstration of a task being performed, robots doing active learning to learn new tasks, robots receiving reinforcement learning from simulations of a task, and robots carrying out deep learning enabled by synthetic data. This growing ability of robots to learn is lessening the need to program robots for specific tasks.
Second, Gupta said, advances in sensing and robot design are enabling robots to perform complex tasks safely and efficiently. These advances include a growing ability of robots to perceive objects with high-resolution, three-dimensional vision; to perceive applied force with high resolution; to sense texture and temperature; to grasp objects with a soft grip; to manipulate multiple objects quickly and with precision; and to interface more effectively with humans. He stated that robots today
are also able to handle various materials, such as fragile objects, granular materials, fluids, non-rigid objects, and microscopic objects, that are of importance in both scientific laboratories and manufacturing. Thanks to such advances, today’s robots are able to interact with their environment in ways that were unthinkable a generation ago, including for sample preparation and moving objects around a laboratory or manufacturing facility.
The bottom line, Gupta said, is that robots can be a useful tool for automating material synthesis and discovery, and he offered four specific conclusions: (1) The advent of human-safe robots is enabling robots to work on ergonomically challenging tasks and amplify human productivity. (2) Physics-informed AI is a key to realizing smart robotic assistants. (3) Advances in sensing are enabling robots to work with a new class of objects. (4) Robots can play an important role in reducing manual labor in materials synthesis and discovery.
Smart Hybrid Machine Tools for Autonomous Materials Characterization and Synthesis
Satish Bukkapatnam of Texas A&M University described how smart hybrid machine tools can be used to overcome challenges in the synthesis, characterization, and reproducibility of bulk-scale materials. Characterizing bulk-scale materials can be time consuming because the geometry, morphology, microstructure, composition, performance, and other characteristics require different tools. This process can take anywhere from 5 to 200 hours, even with the latest innovative instrumentation. Bukkapatnam described work being done at the Texas A&M Engineering Experiment Station on an all-in-one machine tool that combines directed energy deposition addition with subtraction processing, heat treatment, and finishing in a single platform to simultaneously produce components with varied compositions.
The machine was outfitted with 14 different types of sensors, including an accelerometer, a thin-film piezoelectric sensor, a dynamometer, a FlexiForce sensor, a temperature sensor, and a high-speed camera, which collect data on the entire manufacturing process. A key feature is that all of the signals can be aligned to within 1 ms in time and 100 μm in space, making it possible to combine the readings from multiple sensors to get a multi-attribute characterization at one point in time. Bukkapatnam described how his team uses AI to detect morphological defects as a piece is being printed by utilizing a convolution neural network. But, characterizing materials effectively will require developing new kinds unsupervised and active learning approaches.
Panel Discussion
Autonomous Physical Science at the National Institute of Standards and Technology
Gilad Kusne of NIST spoke about three NIST projects that illustrate how autonomous technologies can be used in solid-state physics.
The first was CAMEO, an algorithm that explores new material compositions and learns the relationships among the composition, structure, and material properties. For example, it was used to look at the magnetic properties of iron–gallium–palladium alloys to find the composition that had the highest saturation magnetization. In another trial, CAMEO searched through germanium–antimony–tellurium compositions to discover a new nanocomposite phase change memory material that was superior to the previous best material. The CAMEO exploration and discovery process was 10 times faster than previously possible. And, by adding some physics knowledge into the algorithm and having it run thermodynamics calculations, the NIST group had CAMEO determine a phase diagram by extrapolating beyond the point where it had data. In general, Kusne noted, ML is good at interpolating but terrible at extrapolating; the addition of physics knowledge allowed their system to overcome that limitation.
Then, Kusne touched briefly on two other examples of using autonomous technologies in physics. The first was ANDiE, which was used to accelerate materials characterization using neutron scattering. The other was a Low-Cost Education Platform for Teaching Autonomous Physical Science, a $300 kit for students that allows them to do autonomous science and learn physical laws.
Adaptive Learning Strategies for Efficient Navigation of Search Spaces
Prasanna Balachandran of the University of Virginia discussed how to use adaptive learning strategies to
search databases. He focused on how to move beyond state-of-the-art Bayesian optimization to enable better autonomous searches.
Balachandran is most interested in searches done on “small data”—that is, where the database may be gigabytes in size rather than terabytes of petabytes. His experience indicates that a supervised ML algorithm can achieve improved performance with fewer training data than an unsupervised algorithm, and the search can be facilitated by intelligently sampling the unexplored space in an informed and iterative manner. The following three considerations should be taken to best learn from small data: (1) It is important to quantify prediction uncertainties. (2) One should incorporate domain knowledge as much as possible into the search. (3) Iterative learning should be used for model improvement.
He also emphasized that learning from failures is an important ingredient in the process, and he gave a brief description of work in predicting the compositions of ferroelectric perovskite oxides, where testing showed that the predictions were wrong in a number of cases. Thus, his team retrained the model to include additional information that would correct the errors. It is important to go beyond the black box of ML providing input–output relationships to learn about the causes of those relationships. In this way, relationships identified by ML can point to new scientific understandings.
Coupling of Autonomous Active Learning with Automated High-Throughput Experiments
Michael Thompson of Cornell University spoke about how AI and autonomous active learning have allowed him to carry out automated high-throughput experiments more efficiently and to direct the collection of data in ways that were not possible before.
He sees autonomous processes as just one more tool in his toolkit and that they still require a knowledgeable human experimenter thinking about how to use them efficiently and productively and not simply setting them loose. But autonomous tools have become an essential part of the experimentalist’s toolkit because automation has led to a flood of data that require intelligent autonomy to deal with them. Thompson described experiments at the Cornell High-Energy Synchrotron Source, where he studies the phases that can be produced from varying compositions of materials under varying processing compositions. The experiment involves running a wafer past a laser that heats the materials on the wafer, which quickly cool and settle into their preferred phases, which are then observed via X rays. The process creates about 200 different spectra from different sections from each laser scan, and those spectra need to be analyzed to determine which conditions to create for the next run. Without AI, it would take an experimenter hours to do the analysis. With an autonomous active learning process analyzing the data and deciding on the optimal conditions for the next run, it takes only about 20 seconds, making it possible to do a long string of experiments rapidly.
WORKSHOP RECAP
Hodge provided a recap of the workshop. She commented that the role played by autonomous material discovery differs from field to field, so the distinction between automation and autonomy is important. ML does not have to be a black box but can help elucidate the science involved in various phenomena.
Workforce development will be important to this field, Hodge said, although it is not yet clear exactly which knowledge and skills need to be imparted to people who will be working in the field. It will be crucial to ensure underrepresented minorities are not left behind in being able to take part in the field.
She emphasized that different fields have different tools and will approach design and characterization differently. Because data play a crucial role in enabling ML, thought needs to be given to the best ways to handle and enable access to data. There is much to be done, but ML and AI have the potential to accelerate materials discovery and optimization.
DISCLAIMER This Proceedings of a Workshop—in Brief was prepared by Robert Pool as a factual summary of what occurred at the workshop. The statements made are those of the rapporteur or individual workshop participants and do not necessarily represent the views of all workshop participants; the planning committee; or the National Academies of Sciences, Engineering, and Medicine.
PLANNING COMMITTEE Andrea Maria Hodge (Chair), University of Southern California, Santa Barbara; Thomas Kurfess (NAE) (Vice Chair), Georgia Institute of Technology; David Aspnes (NAS), North Carolina State University; Stefano Curtarolo, Duke University; Robert Hull, Rensselaer Polytechnic Institute; Klavs Jensen (NAE/NAS), Massachusetts Institute of Technology; John J. Koszewnik (NAE), Achates Power, Inc.; Kelly Elizabeth Nygren, Cornell University; Lourdes G. Salamanca-Riba, University of Maryland; Susan B. Sinnott, The Pennsylvania State University; Edwin Thomas (NAE), Texas A&M University–College Station; Haydn N.G. Wadley, University of Virginia.
REVIEWERS To ensure that it meets institutional standards for quality and objectivity, this Proceedings of a Workshop—in Brief was reviewed by Elizabeth J. Opila, NASA Glenn Research Center; Carol Schutte, Air Force Office of Scientific Research (retired); and Michael Thompson, Cornell University. Katiria Ortiz, National Academies of Sciences, Engineering, and Medicine, served as the review coordinator.
STAFF Neeraj Gorkhaly, Study Director; Michelle Schwalbe, Director, National Materials and Manufacturing Board, and Director, Board on Mathematical Sciences and Analytics; Erik B. Svedberg, Scholar; Amisha Jinandra, Associate Program Officer; Joseph Palmer, Senior Project Assistant.
SPONSOR This project was supported by Contract W911NF-17-1-0330 with the U.S. Department of Defense.
For additional information regarding the workshop, visit http://www.nationalacademies.org/event/11-01-2022/autonomous-materials-science-a-dmmi-workshop
SUGGESTED CITATION National Academies of Sciences, Engineering, and Medicine. 2023. Autonomous Materials Discovery and Optimization: Proceedings of a Workshop—in Brief. Washington, DC: The National Academies Press: https://doi.org/10.17226/26989.
