Issue link: https://iconnect007.uberflip.com/i/1493797
MARCH 2023 I SMT007 MAGAZINE 45 that many EMS companies run efficient pro- cesses, such anomalies—what we term "nega- tive data"—can be difficult to obtain. It turns out you can use the principles of gen- erative AI to generate different types of defects, which in turn allows us to create an intelligent system. It is the kind of foundational technol- ogy that has enabled ChatGPT and will enable better AI for many industries, including our industry. Finally, whereas a user cannot necessarily trust the output of ChatGPT, an operator can trust the results of our system by asking it to "explain" how it reached a particular conclu- sion. How easily does Darwin AI integrate into an existing line? Do I need to be fully CFX- capable along my entire line? What's the minimum threshold for bringing you in? We fit right into existing manufacturing lines with inline, standalone, and benchtop solu- tions. e standalone and benchtop versions can be set up at an existing inspection station; the operator can put in the board and check it. With our inline version, boards can travel on a conveyor or sliding rail, so it integrates with existing PCBA machinery. e inline sys- tems are SMEMA compatible. We'll be add- ing CFX compatibility in the future. Although the boards vary, the machinery and workflow are consistent across different suppliers. It was important for us to create a modular solution that could scale across different customers. In the past, golden boards could mess you up because if what you thought was your golden board inadvertently contained flaws, your acceptance criteria would now contain those flaws. Like our discussion of ChatGPT, you could end up building artificial intelligence on inaccurate input information. How do you mitigate that risk? What are the checks and balances? ere are a couple of things. First, we've noticed that some of our clients use two dif- ferent but equally viable components that are acceptable in the context of the production line. ey'll give us multiple golden boards and want the union of that data set to be the source of truth. Second, our system can use a golden board that's passed an ICT/functional test, reducing the risk of a flawed board. Our soware allows the user to remove any com- ponents (e.g., through-hole components, con- nectors) that are not needed at a particular inspection point. e system determines the ground truth extremely quickly and will detect all the com- ponents on the board, but there is still an oper- ator verifying that everything was properly configured. He or she might then run some known flawless boards through the system to ensure functional correctness. We'll never get away from a robust process—setting things up in a way that is significant and trustworthy. With an AI system, the difference is you can do this in less than an aernoon. You have some recognizable customers on your list. How much can you quantify the efficiencies or ROI at the customer site? Our clients are calculating the first-order ben- efits of the system: Is it doing a good job? Is it better or more precise? In terms of rework or eliminating rework cost, what is the benefit? Every time a board runs through our system, we take multiple pictures of it, and we find that, at the technological level, that's excit- ing to thought leaders and executives. ey say things like, "You're making my line more We'll never get away from a robust process— setting things up in a way that is significant and trustworthy.