6
Discussion
Haydn Wadley, University of Virginia, moderated wrap-up discussions at the end of each day of the workshop. At the start of the second day’s wrap-up, he said that he recognized that there have been remarkable achievements, but it is still not possible to go from an optimized design to a manufactured product. He asked attendees to consider the degree to which the benefits topology optimization can bring are constrained by the state of manufacturing today. He and William Paul King, University of Illinois, Urbana-Champaign, also encouraged participants to think carefully about quality control. Can products, including those with novel, unusual geometries—some with thousands of individual geometric features, far beyond the number that existing tools can measure—be manufactured without requiring inspection? If not, what internal inspection techniques, such as data and quality analyses, are needed; can they be automated, and what are the existing gaps?
In addition to gaps and concerns related to knowledge, data, and testing, participants discussed specific struggles in the context of additive manufacturing and multi-physics designs.
STRUGGLES WITH ADDITIVE MANUFACTURING
Joseph DeSimone, Carbon, Inc., agreed that manufacturability remains a key constraint for topology optimization. To fully realize the potential, he said there is a need for a major breakthrough in metals 3D printing akin to that seen in polymers, not just incremental improvements. Investing resources in creating such a breakthrough could lead to truly innovative approaches with large payoffs.
Ryan Watkins, NASA Jet Propulsion Laboratory (JPL), shared that JPL has struggled to infuse additive manufacturing into flight designs. For example, he said, existing technologies have failed to realize the promise of composites, and it is important to understand and fix these issues to expand the design space. Alicia Kim, University of California, San Diego, suggested that integrated computational materials engineering may have faced similar problems that the topology optimization field can learn from.
Jennifer Wolk, Office of Naval Research, suggested that the present culture is another challenge. Today, additive manufacturing is held to the same constraints as conventional manufacturing, despite being very different, because people do not have a realistic understanding of the benefits of topologically optimized designs, she said.
Carlos Levi, University of California, Santa Barbara, pointed out that another cultural challenge is that design and manufacturing departments have traditionally worked separately, but now, because of materials development, that is no longer the case. Finding new alloys that are more manufacturable, which has helped in the past, could invite better collaboration.
Noting that 3D printing is still a new technology, Reinhard Radermacher, University of Maryland, suggested that it should be integrated with existing parts to open up new opportunities, instead of merely replacing traditional systems, in order to transcend today’s constraints. For example, 3D technology could one day enable a heat exchanger to be melded with a fin or for an entire system to be composed of a single additively manufactured component.
SOLVING NARROW PROBLEMS WHILE ADVANCING FUNDAMENTAL UNDERSTANDING
Hardik Kabaria, Carbon, Inc., stressed the importance of solving specific design problems. Instead of trying to achieve surface perfection, for example, perhaps manufacturing constraints should be accepted and designed for instead. The incremental successes made there could then be applied more holistically to help additive manufacturing mature, he suggested. Jonathan Berger, Nama Development LLC, added that it is also important to look for opportunities where the inherent properties of additively manufactured parts, such as surface roughness, can be considered to be features rather than defects.
Another participant stressed the importance of asking the right questions and being wary of imposing constraints or expectations on new technology. Developing specialized parts for unique problems teaches people how to make things better, the participant said, but that is still a long way from manufacturing profitable consumer products. Trial-and-error experimentation is possible at the research level, but additive manufacturing alone cannot solve every problem. Radermacher
agreed, noting that in his experience, companies are very cost conscious, but if they can see performance benefits, they are more open to new ideas.
Ned Thomas, Rice University, wondered if the field could be advanced through a focused Grand Challenge by the Defense Advanced Research Projects Agency, similar to the one for advancing self-driving cars. Milton added that scaling, especially for multiscale structures, is a challenge at the small scale, with defects and other worries, but topology optimization could be helpful at the large scale.
Kim pointed out that performing topology optimizations is an important part of a fundamental understanding of science, in addition to informing designing and building. She noted that metal-based additive manufacturing is moving into higher technology readiness levels, but she suggested that what is needed is more support for fundamental research to enhance understanding of how characteristics like surface roughness and porosity can be predicted and managed in a multi-physics, multi-materials environment.
Kimberly Saviers, United Technologies Research Center, suggested that pursuing global optima and minima may not be a productive path, and Manoj Kolel-Veetil, Naval Research Laboratory, suggested incorporating artificial intelligence and machine learning into this space, which works well even with small amounts of data and as long as the laws of physics are still respected. There are so many fundamental questions that need to be better understood, he continued, that this method could help advance progress.
TESTING GAPS
Several participants noted key gaps with regard to the ability to test topology-optimized designs. Christopher M. Spadaccini, Lawrence Livermore National Laboratory (LLNL), stressed the need for better nondestructive evaluative methods, which are currently slow, expensive, and geometrically inaccurate. Another gap, he added, is that qualifying non-additively manufactured parts currently requires expensive metrology machines. To overcome some of these limitations, he suggested, it may be possible to learn from semiconductor design rules to design and optimize rapid, measurable, built-in test and inspection structures.
Rebecca Dylla-Spears, LLNL, suggested that instead of testing certain geometric features, perhaps fiducials, critical to the design but judicially chosen, could be built to flag errors and identify potential damage. Edwin L. Thomas, Rice University, suggested that flow could be tested through an increasingly specific hierarchy, creating a testing gradient that could accurately measure fluid or air flow. Kabaria agreed that flow transfer testing could measure quality and noted that it could also be useful at Carbon to catch and clean excess viscous resin. Wadley agreed, noting that the process could also work well for lattice structures. Katherine T. Faber, California
Institute of Technology, wondered if it could also identify microstructural issues, which tend to be very difficult to detect.
Dianne Chong, Boeing Research and Technology (retired), noted that in her experience, with large parts, not every detail is critical and needs qualification. Tailoring parts, and understanding the physics in certain segments, is important, however. At the macro level, technological tools can consolidate and test multiple parts at once, but she stressed that it is important to include designers and structural analysts in the entire process to establish qualifications and create confidence in the designs, especially for new technology.
KNOWLEDGE AND DATA GAPS
Ole Sigmund, DTU Technical University of Denmark, stated that optimization is only possible if it is also possible to quantify quantifiability, and King replied that a key gap lies in determining how to quantify this. Once this question is answered, it will be possible to create algorithms for scoring parts, although the next challenge will be scoring a part with thousands of geometric features, a necessary step before optimization and one that will require more new tools, he added.
Noting that most 3D-printing processes are actually layered 2D printing, albeit with multiple geometries and in situ monitoring, a participant suggested that it may be possible to harness that monitoring data, quantifying it for use in topology optimization, and enable feature adjustments to avoid common defects and improve outcomes. King agreed, pointing out that the part during the process is very different from the part after the process.
Several participants raised questions about the data that are generated during processing. One participant mentioned that “born qualified” capabilities include a part’s entire data set, known as its “digital twin,” to speed up testing, maintenance, and reprinting. These data are also important to identify failure points, the participant added. Spadaccini agreed that saving data is critical, yet in his experience, many production partners prefer not to, in order to protect themselves against being held liable for future failures. Chong agreed, adding that companies may also fear data theft.
MULTI-PHYSICS CHALLENGES
King reiterated the need to better understand and address the complexities that arise when working with multi-physics packages, especially when interfacing with optimizers. For example, for mechanical structures that interact with fluid flow, there are many nonlinearities and viscous effects that add optimization challenges. Sigmund added that in the case of fluids, turbulence is also a challenge, as it is extremely complex to model, with inconsistent and nonstandardized
sensitivity analyses. To overcome that hurdle, analyses are built on top of a physics package followed by automatic differentiation, which is easier than building them within it but may create inaccurate outcomes. Claus Pedersen, Dassault Systèmes Simulia Corp, agreed, adding that reprogramming simulations to accommodate nonlinearities also creates challenges. The company has incorporated nonlinearities into its optimization software, he said, but it has been difficult and sometimes requires various optimization attempts by the user.
King asked participants to comment on whether they have been able to use topology optimization on multi-physics packages. Wadley suggested that the work of Xiaoyu (Rayne) Zheng, University of California, Los Angeles, with piezoelectrics was a promising start to address multi-physics interplays. Sigmund agreed, and he also recommended COMSOL, with its automated sensitivity analysis, to address multi-physics issues. However, he noted, it is not yet able to optimize all physics combinations, such as aerodynamics with elasticity. Creating methods to combine multiple physical modeling packages represents both a challenge and an opportunity, he said.
Zheng noted that some field-responsive materials have dynamic properties that can be tunable to their environments, but it is not yet clear if this quality can be incorporated into topology optimizations. Sigmund agreed, noting that if modeling is possible in the time domain, then those qualities could be optimized. He reiterated his point that if data can be modeled, they can be optimized.
Another participant pointed out that heat transfer has physics gaps, such as analysis tools to study two-phase flow. For flow optimization, the process needs to combine multiple discrete elements, and, again, there are no existing tools to do it quickly.
Angus Kingon, Brown University, replied that he sees two alternate approaches: (1) include multi-physics in an integral way or (2) stitch the processes together serially. Each approach has upsides and downsides, requiring tradeoffs, for example, in the length of time it takes or in the user and modeler outcomes. Sigmund concurred, noting that while combining modeling packages is very complex, unfortunately there is no single package that fully satisfies all physical modeling needs.
William Benard, Army Research Laboratory, pointed out that the real world is multi-physics, and it is difficult to design for something so complex. Optimizing in just one dimension, or two at most, creates new and unexpected failure modes. It may be possible to address these failures through the creation of workflows for automated designs, incorporating topology optimization, he suggested. For example, if the Army wanted to create a new ground vehicle, a workflow should cover all the factors required, such as temperature range or chemical environments. There may be a high setup cost to establishing such a detailed workflow, but it would ultimately be an advantage, for example, by reducing the time and expense of testing a new material or adjusting to a new threat, he said.