EC-25.5 Reliability: Wear-Out, Burn-In and Lifetime

The standard treatment of semiconductor reliability: the hazard curve, burn-in, wear-out mechanisms, acceleration models, life distributions and qualification, September 2026

What this is and why it exists

A part that passes every test on the day it is made can still fail after two years.

The mechanisms that cause this are physical, understood, and predictable in aggregate even though no individual failure can be predicted. That is what makes reliability a statistical subject rather than a matter of quality control.

The whole of qualification rests on acceleration: stress a sample harder than the field will, and use a model to convert the result. That model is where the argument usually is, and understanding its assumptions separates a qualification from a ritual.

The vocabulary

  • Hazard rate — the probability of failing now, given survival so far.
  • Infant mortality — early failures caused by manufacturing weaknesses.
  • Useful life — the long middle period with a roughly constant hazard rate.
  • Wear-out — the rising hazard rate at the end of life.
  • Burn-in — stressing parts briefly to remove the weak ones before shipping.
  • Oxide breakdown — the gate insulator failing after prolonged electric stress.
  • Electromigration — metal atoms moved by current, eventually opening a connection.
  • Activation energy — the constant relating failure rate to temperature.
  • Qualification — the stress sequence a new product must survive.

The mental model

Failure rate over life has three regions. It is high at first, because manufacturing weaknesses fail quickly. It is low and roughly constant for a long period. It rises again at the end, as wear-out mechanisms accumulate. Quoting one failure rate for a product ignores which region it is in, which is the most common misuse of a reliability figure.

Burn-in exploits the first region. Run parts hot and biased for a period, and the marginal ones fail in the factory instead of at a customer. It costs time, equipment and a little life from every part that survives, so it is used where a field failure is expensive.

The wear-out mechanisms are worth naming, because each responds to different stress. Oxide breakdown responds to voltage. Hot-carrier damage responds to current and to switching. Bias-temperature instability responds to voltage and temperature together. Electromigration responds to current density and temperature. Which dominates depends on the technology and on how the part is used.

Acceleration is the bridge to a prediction. Stress a sample at high temperature and voltage, measure when it fails, and convert to field conditions using a model. The usual model relates rate to temperature through an activation energy. The prediction is only as good as that constant, which is measured on the technology rather than derived from theory.

Life distributions matter more than averages. Failures cluster in a distribution with a characteristic shape, and what a customer cares about is the fraction failing by a given date. A mean lifetime tells you almost nothing about that fraction, and a large mean can hide an early tail.

Finally, qualification assembles all of this. Temperature cycling, humidity, high-temperature operating life and mechanical stress, in a defined sequence with a defined sample size. Passing it is a condition of selling into most markets.

What you should now be able to explain or do

  • Describe the three regions of the hazard curve and their causes.
  • Say what burn-in removes and what it costs.
  • Name the main wear-out mechanisms and the stress each responds to.
  • Convert a stress test result into a field lifetime using an acceleration model.
  • Explain why a life distribution says more than a mean lifetime.
  • Outline what a qualification sequence contains.

Check yourself

The weakest parts, which would otherwise fail early at a customer. It moves infant mortality into the factory, at the cost of some life from every part.

Current density, together with temperature. Metal atoms are moved along by the current, and heat increases how easily they move.

Because the whole extrapolation depends on it, and it is measured rather than derived. An error there scales the predicted lifetime by a large factor.

Because customers care about the fraction failing by a date. A distribution with a long tail and an early shoulder can share a mean with a much better one.

Go deeper

Back to Reliability: Wear-Out, Burn-In and Lifetime: work through the checklist