Issue link: https://iconnect007.uberflip.com/i/1172746
70 SMT007 MAGAZINE I OCTOBER 2019 looking for very specific outliers of an expected outcome. Of course, AI can be used to the benefit of reliability with predictive modeling, but all of the statistics needed for the input of that modeling come from hands-on analysis. I was recently part of an iNEMI project that looked at the cleanliness of quad-flat no-leads (QFN) components after assembly with the idea that if you know how much ionic residue is present, you can predict the risk of that part failing in a normal field service environment. That was done with a consortium of companies around the world using ion chromatography, surface insulation resistance (SIR), cross-section, and scanning electron microscopy (SEM) imag - ing to create a very large dataset that was ana- lyzed within an inch of its life to get a defini- tive answer on the inherent risk of using BTCs. I won't spoil the ending, but the answer is like most cleanliness-related issues, it depends. (I think I now owe Doug Pauls a quarter.) It depends on what type of testing you do and what your sample matrix is comprised of as well as the end-use environment and expected life of that product in that atmosphere. In gen - eral, you need to perform ionic cleanliness eval- uation using ion chromatography and elevated heat and humidity test under power at a mini- mum. Currently, there is no single machine or piece of software that can do both tests. The ability to recognize residues was done with several types of analysis and the experience of engineers and operators who did not rely on a single machine that lacked the tribal knowl - edge of that team. When I think of AI today, I think of things like autonomous vehicles; if you do a quick online search, you will find several videos of people asleep at the wheel while going 60+ miles per hour on the highway. Does anyone think this is a good idea? Sure, we would all like to catch a quick nap on the way to work, but it seems like an abuse of a system designed to help with safety. I liken that to relying on only what one piece of equipment tells you about cleanliness because it's in a drawing specification that has been around much lon - ger than it should have been. In the videos with snoozing drivers, there isn't a horrific crash at the end, so a lot of peo- ple think, "Well, it worked for that person, so no harm, no foul." But to me, it's just a matter of time before it doesn't end well. This is how I relate it to reliability; we get comfortable when things are going well, and there isn't a mas - sive rash of field failures. The assumption is that your process is on cruise control (no pun intended), and there can be a tendency to back off due diligence testing that was done to qual - ify the process in the first place. Without peri- odical testing in a manner, like what was done for first article inspection, there could be creep- ing variability not being detected that could eventually steer you into the ditch (pun fully intended). Eventually, I assume that AI will replace us all, and I, for one, welcome our robot over- lords. But until then, we still need to put in the work to ensure that we are doing a hands- on inspection when a solder joint is in ques- tion or when the reflow oven doesn't seem to be running just right. When used in conjunc- tion with actual experience and intelligence, AI can certainly help increase reliability, but we aren't there yet. Let's ensure that we are still relying on humans to do most of the criti- cal thinking. SMT007 Eric Camden is a lead investigator at Foresite Inc. To read past columns or contact Camden, click here. This is how I relate it to reliability; we get comfortable when things are going well, and there isn't a massive rash of field failures.