July/August_AMP_Digital

A D V A N C E D M A T E R I A L S & P R O C E S S E S | J U L Y / A U G U S T 2 0 2 0 3 3 create a defect, well, that would be dy- namite. And I think doing that without some kind of AI is just not going to be realistic. How might AI techniques be applied to additive manufacturing?  Holm: Additive manufacturing is unique in that it gives us terabytes of data from the in-process monitoring. And that’s different from the other di- aled-in processes we’re used to. This is unique data for every single build and it is noisy and it is huge, but there’s defi- nitely signal there. You could even see it when you’re watching or listening to the machine build. And the only kinds of tools we have to deal with that type and quality of data are the deep learn- ing, big data tools of AI. And so I think the big win for AI would be the ability to do in-process monitoring, especially, in-process control for outcome. That’s the holy grail. Williams: The other thing to con- sider is, if you have sensors for individ- ual process variables, then you could use AI to rank the importance of those process variables, so when the day is done, you knew which ones you really had to monitor and which ones would be much less important. And that can change depending on the component and the intended application. What are challenges or barriers to using AI in additive manufacturing? Herderick: I had a project where my students tried five different topolo- gy optimization software programs, and they got five different answers. Most of those are not actually materials-based. So there’s no software out there where you can actually go in and give it a fa- tigue limit, or a corrosion limit, or any of the critical design to no-fail properties that are actually used in real hardware. There’s no digital tool for that. We use semi-empirical modeling, but it’s not like you can put a fatigue limit in a crit- ical area into your topology optimiza- tion and have it take out your material, right? That’s still an iterative process. So I think that’s a big opportunity for ex- tending our use of this data. Holm: I agree. It’s all fed by just the sheer size and diversity of the data we can get. All of a sudden, we can in- tegrate experience and create a knowl- edge base that we can sustain and continue to use, to continuously im- prove, and I think that’s exciting. Williams: But we need to empha- size that spending money on design data and design curves is a waste of time and money until you can demon- strate that the process is under control. Herderick : If we’re trying to put something out there in terms of a grand challenge for the materials communi- ty, one idea is understanding the nature of defects created during printing, the effects of these defects on properties, and how to inspect for them before in- stallation. I see a lot of work being done on the indications that you get from in situ monitoring. What I don’t see is ty- ing those to actual creation of defects. And because of the digital twin value chain, from in-process monitoring, the first time quality inspection, like in an X-ray CT, you would do in a factory with, for example, a turbine blade. Detailed first-time inspections to define a pro- cess window is necessary certification coming out of the factory. What is the one thing that AI could bring to the AM process that would be the most beneficial? Herderick: It would be verification and validation of a well-controlled pro- cess and tied to that would be effects of defects. Williams: I agree with that, Ed. But embedded in what you just said is the assumption that there’s agreement on what a well-controlled process is, and I don’t think we’re there yet. Until we can understand the variation associat- ed with ordinary fluctuations in process and can define the bounds of those vari- ations, then you really can’t statistically analyze the data because it’s not from a consistent source. Holm: I would add, as the AI per- son here, what additive practitioners can do that’s the most valuable, is to keep the data. More data is better data. Data from more sources is better. And we need to not only keep the data but aggregate the data and help each oth- er in scaling up this new technology. ~AM&P