Modern industrial and mechanical engineering teams have enabled the production of nearly every good at population scale. Supporting these massive operations are control practices to ensure that each item leaving the facility exceeds the expectations of regulators and consumers. But in mass manufacturing, excursions from quality control (QC) limits are an inevitability. The question is not how to avoid QC failures altogether, but how to build a visionary system to predict QC issues before they even arise.

One such way to shift from reaction to prediction is through statistical process control (SPC), wherein limits to the QC standards can be calculated rigorously and batch quality can be continuously monitored. Sarah Gilyard, Senior Product Manager at JMP, explains that “variation is a part of every manufacturing process, but it’s something you want to reduce as much as you can. Monitoring the process and tracking its performance is key.” In doing so, process engineers can differentiate common-cause variation, the inherent natural range of outputs, from special-cause variation, which is caused by a defect in the process that necessitates remediation.

Having spent 16 years in fab processing and equipment engineering roles prior to joining JMP, Gilyard experienced the semiconductor industry’s transition to SPC directly. Under the previous paradigm – and in a sector where precision in manufacturing is paramount and single grains of dust violate QC standards – Gilyard says the team “would identify a problem at the end of the wafer manufacturing process and then react to that. We would search for what the problem was.” While generally effective, this method was time-consuming and resource-intensive, she says. She continues that the team moved toward “a robust SPC program where we could track processes and monitor variation in real time. We solved a lot of problems before they happened.”

The major change was to stop evaluating the success or failure of the entire process at the end, at which point the exact cause of the issue may not be clear. Instead, real-time monitoring of variance using statistical techniques allowed for instant detection of QC limit excursions, while simultaneously providing pinpoint evidence for what the failure mode was. With all the benefits of SPC, then, Gilyard asks, “Why aren’t more companies thinking this way? In the presence of vast quantities of data, why do some companies still revert to firefighting?”

The answer lies in the resources needed to establish a robust SPC program, as well as the mindset shift required on multiple levels of the corporation. Gilyard notes that “it’s not enough to simply purchase an SPC software package,” instead explaining that “companies need to invest training and time not only into training their workforces about what SPC is, and how to interpret control limits, but also cultivating that mindset where they want to solve problems before they happen.”

Having that viewpoint widely adopted requires intentional framing and resource investment, especially when engineers may not have been trained in statistical methods. But the result of investment into a new companywide paradigm can ensure a more proactive workforce, where the prevailing idea is not only fixing what is already broken, but that “if it’s trending toward broken, that is a problem too,” Gilyard explains. However, solely changing the mindset of the engineers in the fab can’t prompt an enterprise mindset shift. For that, Gilyard says, company leadership must learn to balance resource investment with risk profiles and see that SPC bears a worthwhile upfront cost in the long run. To help frame that balance, Gilyard and the team hosted “SPC meetings where we would talk about the controls for certain processes and what our process capabilities were looking like.”

Friction can be part of the shift, she notes, because “there is a proper way to calculate SPC limits,” whereas management may be accustomed to simply setting a desired QC value. But as everyone gets desensitized to the new paradigm, Gilyard describes a gratifying conclusion: “If you understand your risk profile, then you can set your controls appropriately based on that risk profile.” Doing so successfully bridges the perceived gap between resource allocation and risk management. In the process, company leadership and fab engineers can align their mindsets while calculating control limits that set them up for predictive success.

Aiding with prediction are the AI tools currently available at enterprise scale. Now more than ever, highly specialized tools can be incorporated into a data-driven workflow. Gilyard advocates for targeted usage of AI, without an overreliance on it. With the proper guardrails, AI tools have the potential to assist engineers in predicting special-cause variation before it corrupts a batch; each time SPC catches an issue before it’s too late, it builds trust in leadership that a statistical mindset is worthwhile.

Gilyard concludes that a misconception in SPC is that “it’s all about statistics.” Instead, she argues that SPC is much more about “deciding all the things before you perform the calculations and implement the statistics. It’s the logistical piece of it.” But foregoing SPC, she argues, is worse, because “you can quantify how many excursions you could have saved if you had known about them earlier. In almost every situation, you’ll find that allowing excursions to happen without detecting them sooner – and allowing bad product to get out the door – tallies up to a sum of money that’s greater than what you’d spend on implementing an SPC program.” Only by constructing a data-driven workflow can companies make the full shift to AI-ready smart manufacturing and reap the benefits therein.