End-of-Life Analysis: Application of Degradation Models

The day the CPK report came out, the entire room fell silent for three seconds

I remember several years ago, when we launched a new product, everyone was fixated on the yield rate. At that time, the CPK report from the test station showed that a key parameter's CPK was a mere 1.08, with DPMO as high as 6210! The PM immediately jumped up and asked, "How long can this product last? Will it fail shortly after customers install it?" Everyone exchanged glances, because we hadn't conducted a complete lifetime test; we just relied on experience and said, "It should be fine." The result? Three months later, customer complaints started pouring in; the product indeed didn't last long. From then on, we truly began to prioritize "End-of-Life (EOL) analysis." To be honest, before that, we only thought about putting out fires when they were already burning fiercely.

Where the problem lies

Simply put, EOL analysis is about understanding how long your product can last before it "starts" to have problems. This doesn't mean complete failure, but rather when its performance begins to degrade noticeably, reaching a point unacceptable to customers. We often say "death is imminent," but in a semiconductor fab, we're actually more concerned with the point where the product becomes "critically ill."

Previously, we only looked at "yield," meaning whether the product was good when it left the factory. But good doesn't mean good forever, right? Your phone's battery is great when new, but after a year, it starts draining quickly. It's the same for products: under the stress factors of time, temperature, and voltage, their electrical performance will slowly drift, eventually potentially drifting out of spec. EOL analysis aims to establish a "degradation model" to predict how fast this drift occurs and when it will "cross the boundary" (go out of spec).

How it's done in practice

The most direct approach is to take a batch of products and "stress them to failure" under accelerated aging conditions. For example, place products in a 125°C oven with continuous power applied, then periodically take them out to measure their key electrical parameters. For instance, we have an analog circuit where the output voltage offset is a critical parameter. We assume the offset voltage of a new product is within 10mV, but the specification requires it to be less than 20mV.

1. Collect Data: We take 50 units, run them at 125°C and rated voltage for 1000 hours, measuring the offset voltage every 100 hours.
2. Establish Model: You will find that as time increases, this offset voltage gradually drifts upwards. If you plot these data points, you'll typically see a trend line, which could be linear or exponential. This trend line is your degradation model.
3. Predict Lifetime: Based on this model, you can predict approximately when the product's offset voltage will "drift past" the 20mV specification limit. Let's assume we calculate that it will drift past at 2500 hours.
4. Convert to Real Lifetime: Don't forget, this is accelerated lifetime, which must be converted to lifetime under real usage conditions using the Arrhenius Equation or other acceleration factor models. For example, if our acceleration factor is 10x, then an accelerated lifetime of 2500 hours means that at the customer's end, it will take approximately 25000 hours (about 2.8 years) before it starts to degrade to the point of affecting functionality.

So the key is, you shouldn't just look at "whether it's good now," but also "whether it will fail in the future."

The most common pitfalls

The most common pitfall I've encountered is that people often only perform "failure analysis" instead of "degradation analysis." This means analyzing why something failed *after* it has already broken, instead of predicting when it will start to "deteriorate" *before* it fails.

1. Only test initial yield: Many teams only care about the product's yield rate when it first leaves the factory, believing that once it passes testing, everything is fine. As a result, customers only discover after using it for a while that it's not durable at all.
2. Choosing the wrong monitoring parameter: Sometimes engineers choose a parameter to monitor that is not very relevant to product lifetime degradation. The result is that after extensive testing, the data looks great, but the product still fails when it's "supposed to." Simply put, you need to find the key indicator that truly "deteriorates."
3. Misusing acceleration models: The worst part is, after finally conducting accelerated testing, people arbitrarily apply an acceleration factor when calculating the real lifetime. Frankly speaking, the acceleration factor varies for each product and each failure mode; it cannot be generalized. An unvalidated acceleration factor means the calculated lifetime is just self-deception.

One thing you can do today

Identify the most critical electrical parameter for your product and start monitoring its trend under accelerated aging.