A Vision for Machine Learning in Optical Inspection

Estimated reading time: 2 minutes

DarwinAI is developing machine learning and artificial intelligence solutions to improve AOI inspection results. Their work is unique because it’s generally independent of the hardware used in the EMS line. It represents, as much as possible in our industry, a pure-play solution using machine learning and AI constructs. 

Earlier this year, the I-Connect007 staff spoke with DarwinAI CEO Sheldon Fernandez about artificial intelligence and electronics manufacturing. In this follow-up, Sheldon discussed what AI and machine learning are, what they aren’t, and how this work fits into the manufacturing environment of today.

Nolan Johnson: Where does artificial intelligence fit in an industrial environment like electronics manufacturing? 

Sheldon Fernandez: We’ve focused on visual inspection. The step-change capabilities of artificial intelligence to do such inspection is becoming a significant presence in the industry. We focus on the final third of the workflow—back-end post-SMT assembly—which represents a fairly underserved aspect of the PCBA process. It’s also an obvious area where you see immediate productivity improvements through deep learning and artificial intelligence. 

It would be irresponsible not to mention, of course, the explosion in the AI field with generative AI, which is what ChatGPT is based on. This is when AI generates assets that are beneficial to human beings in any context: natural language, images, generating computer code, and scripting code. You already see this for coding but think about the potential of generative AI for the PCB designers. I imagine those who create design tools—as well as the designers—may leverage this technology in productive ways. I want to be clear: It won’t be perfect. But if it gets me 70% of the way there, I can do the rest. Those are the two areas that immediately jump to mind as someone in the AI field.

Johnson: On the design side, we've had auto routers for a long time. In semiconductor, as early as 25–30 years ago, there were IC chip compiler tools. Would those tools be called AI?

Fernandez: They might be. When we say AI, what are we talking about? It's a term that now has come to mean anything that generates a useful asset or insight. Practitioners in the AI field itself typically take AI to mean second or third wave machine learning and deep learning: the facet of artificial intelligence that leverages complex models that learn by analyzing vast amounts of data. No doubt that some of the tools you're describing did a lot of this work. Now, I'm not an expert on the design side, but what is the step change in productivity? In software development, for example, there have always been tools which will help you debug quicker and help with code generation. But there hasn't been a tool where I can say, “Write an algorithm to sort one million numbers with these properties in this time with this algorithmic specificity,” which delivers Python, C-Sharp, or Java code. The ease with which someone who's non-technical can get an output like that is quite remarkable. It's that element of sophistication when I refer to AI.

To read this entire conversation, which appeared in the July 2023 issue of SMT007 Magazine, click here.