2

Data, Statistics, and Analytics for Additive Manufacturing in the National Laboratories

The workshop’s first panel, moderated by planning committee member Maria Emelianenko, George Mason University, had four speakers from the four sponsoring national laboratories—Los Alamos, Idaho, Lawrence Livermore, and Sandia—providing context and background on additive manufacturing (AM) from the perspective of the national laboratories. A major reason that the laboratories are interested in AM is its potential for use in maintaining the nation’s nuclear stockpile, which requires working with manufacturing techniques and materials that were developed during the Cold War in the 1960s through 1980s and looking at how to enhance the stockpile in the future by taking advantage of new tools and manufacturing techniques.

MANUFACTURING ISSUES

John S. Carpenter, Los Alamos National Laboratory, kicked off the panel by talking about the manufacturing aspects of maintaining the nation’s nuclear arsenal. The National Nuclear Security Administration (NNSA), which has the overall responsibility for maintaining the arsenal, works with facilities around the country, including the Savannah River National Laboratory, Y-12, Pantex, and the National Security Campus in Kansas City as well as Los Alamos, Lawrence Livermore, and Sandia, all of which are part of the Department of Energy (DOE).

The DOE laboratories focus on design and manufacturing missions for making small numbers—often tens to hundreds—of highly specialized

parts that are intended for use in extreme environments. The parts often contain uncommon materials, and there should be flexibility in how the parts are manufactured. The national laboratories’ purpose in funding the workshop was to identify collaborations and tools that could help them perform their mission better, Carpenter said.

One of the key issues that NNSA wants to address is finding ways to maintain nuclear weapons built decades ago, when the materials and manufacturing methods from that era are increasingly unavailable. This is particularly difficult because the qualification metrics for the various parts were process-based, so if a 50-year-old machine has broken down or if a particular material is no longer available, it becomes necessary to not only find a new way to produce the part but also to qualify the process.

Carpenter sketched out a four-pronged approach that the national laboratories are taking to address this challenge:

• Anticipate materials issues before they reach the critical path;
• Deploy advanced techniques to reduce materials development cycles, including artificial intelligence and machine learning (AI/ML) and digital manufacturing methods;
• Focus on manufacturability through modern, digital manufacturing methods; and
• Integrate with emerging digital transformation, including AI/ML approaches.

The goal, he said, is to achieve a robust, responsive ability to discover, scale up, and deploy manufacturable and qualifiable new and replacement materials.

AI/ML has a huge potential to revolutionize the entire nuclear weapon life cycle, Carpenter said. It can be used in material discovery, design optimization, manufacturing and certification, and deployment and surveillance. Also, using a digital thread there may provide a seamless flow of digital product information across the product life cycle, tying everything together effectively.

MATERIALS DESIGN

Allen Roach, Idaho National Laboratory, explained that, unlike the other three national laboratories represented on the panel, which deal with nuclear weapons issues, the Idaho laboratory deals with nuclear power issues. The extreme environments involved in working with nuclear power drive the laboratory’s need for new materials that can operate in a wide range of environments.

Relating the AM processing parameters to the properties of the materials produced and predicting the materials’ performance in various extreme environments requires experimental capabilities, physics-based modeling, AI, and data analytics. The Idaho laboratory developed a 5-year strategy for developing AM materials for use in extreme environments that includes plans for material design, material characterization and testing, manufacturing process development, and data analytics integrated with modeling and simulation for material design and selection. The particular materials of interest are nuclear fuels, lightweight materials, and advanced survivability/protective materials.

The laboratory created the Digital Innovation Center of Excellence (DICE) as a “virtual center to formalize and coordinate digital engineering, digital twinning, and digital transformation activities across next-generation energy systems.” It is in part a repository for the storage of digital twins developed for such things as materials and parts. But these models will be accompanied in the repository by information on multiscale modeling frameworks for process-informed material design, process optimization and control, and testing and characterization.

Roach concluded by emphasizing that the qualification and certification of additively manufactured components begins with quantifying the effects of process variability and understanding what causes that variability.

EXAMPLES OF PROJECTS AT LAWRENCE LIVERMORE NATIONAL LABORATORY

Chris Spadaccini, Lawrence Livermore National Laboratory (LLNL), said that the laboratory is carrying out multiple AM-related projects. In one, test structures are built using lasers that fuse powder into a desired shape, then the structures are scanned, and the data from those scans are used to train a convolutional neural network that identifies the optimal parameters. This approach leads to improvements in the structures of the parts, he said.

A second approach involves a custom-built printhead with a mixing nozzle that prints with a mix of different materials, changing the composition as it prints. The printhead can be moved with high precision along five different axes to create complex shapes. The machine also has various observational instruments attached: a CMOS camera, a confocal probe, and a force feedback sensor that measures the pressure with which the materials are pushed through the nozzle (Figure 2-1). These instruments can produce a variety of observational data as a structure is being laid down, and those data can be reduced to useful information, which can then be used to make decisions about the operation of the machine. The

data can also be used to detect defects or get other structural information about the components being produced.

LLNL has also been doing a lot with digital twins, Spadaccini said. For example, a scalable pipeline is being developed to test mechanical performance via digital twins. It uses a process-level digital twin and a part-level digital twin to simulate mechanical response and assess the performance of a part digitally.

WHAT IS HAPPENING OUTSIDE THE NATIONAL LABORATORIES?

Tyler LeBrun, Sandia National Laboratories, spoke briefly about the sorts of additive manufacturing work being done outside the national laboratories. It is important to recognize and acknowledge that work, he said, because it offers sources of additional resources as well as potential collaboration partners.

Many of the companies putting the most work into additive engineering are in a few specific industries that are heavily regulated, such as commercial aviation and medical devices, with investment coming from a variety of private and governmental sources. LeBrun said that for industries representing the greatest share of investment, the data and statistics resemble the brute-force methods previously used to develop new materials. In particular, the data usage and analytics are not fast, efficient, or affordable. However, he suggested, it may be possible to use such things as in situ process monitoring to accelerate the work.

The key issue, he said, is how to turn data into information that enables intelligence. The old ways of reliance on only data are expensive and slow. The goal should be to use intelligent process and tool qualification to enable the rapid qualification of more diverse products. The alternative is to be forever trapped using bulky datasets to validate and verify.

DISCUSSION

The question-and-answer period began with a discussion of the types of datasets that already exist and which of those datasets are being shared. Some owners of datasets share them with collaborators, but many of the datasets that are available for sharing are generated in research and development laboratories and not production environments. One platform that is completely open is DICE, which is open source and available to everyone.

In response to a question about how to use data to power predictive models used for such things as online monitoring and repair, Spadaccini said that a key is to use reduced-order models that can get an approximate answer quickly. There were some concerns expressed about what happens when parts are repaired or modified during the AM process in response to the monitoring, because every part could end up being made differently. That may be acceptable, however, because there is variability present in traditional manufacturing as well; the key question is what types of variability are concerning. These are the sorts of things that need to be considered during the design process. For instance, designing something that requires extremely tight tolerances may not be workable.

Planning committee member Wei Chen, Northwestern University, commented that digital twins need to be more than just a virtual simulation of the existing system; they also need to bring in such things as real-time information from sensors about the part being manufactured, as well as make decisions about controls and processing conditions, which makes it possible to create an autonomous system. Has anyone done a cost–benefit analysis of introducing such autonomous control? There has been no such quantitative analysis, Spadaccini said, but he and other respondents believed that such autonomous control could be very valuable, depending on the specific application.

Multiple speakers suggested that one area that has not received enough attention in AM is post-processing, because most of the emphasis has been on the actual fabrication of materials and components. Carpenter, for instance, said that there is a “huge opportunity space” in optimizing post-processing and understanding how that will affect a part’s properties and performance, while LeBrun added that “we haven’t been doing a good enough job yet to further the downstream elements of the processing

of parts, and much of the emphasis has been on the upstream side of the actual fabrication of the material.” The bottom line is that post-processing can play a large role in the properties of an AM product, and it deserves more investigation.

In response to a question, Carpenter said that one of the major advantages of AM is the ability it offers to control the microstructure of materials, which provides benefits in certain environments and conditions. To take advantage of this capability, it is important to develop modeling capabilities and learn more about the performance of different microstructures in various situations.

One audience member asked for guidance around what data to keep and what data to store during the data collection process. Roach and Spadaccini agreed that while it is currently possible to generate tremendous amounts of data regarding the AM process, it is still not clear which data are most important. It is not always clear what data are needed, and not all data are created equal. Roach remarked that it is also not just about the right data, but getting enough of the right data. Spadaccini added that there is a question of what can actually be measured. “We have a lot of thinking to do on that question,” Spadaccini said.

In response to a question from planning committee chair Thomas R. Kurfess, Georgia Institute of Technology, LeBrun and Spadaccini noted that AM is being held to a higher standard than, say, traditional casting because today’s measurement tools make it possible to see all of the flaws in an AM component; similar flaws may be present in a cast piece, but experience has shown how much of a safety margin should be allowed. In that sense, LeBrun said, AM is a capable tool that is “the victim of its time.”