In short: Early-life (infant-mortality) failures are the steep, falling front edge of the bathtub curve caused by latent manufacturing defects, and burn-in deliberately over-stresses parts under elevated voltage and temperature so those weak devices fail in the factory instead of in your customer's hands.
Every population of integrated circuits carries a small fraction of parts that look perfect at wafer sort and final test but are quietly defective. They power on, they pass functional vectors, they ship — and then they die after a few hours or a few weeks in the field. Understanding why these parts exist, how they are modeled by the bathtub curve, and how burn-in screens them out is core post-silicon reliability knowledge that shows up in nearly every product-engineering and reliability interview.
The bathtub curve: three failure regimes
Plot the instantaneous failure rate (the hazard rate, usually in FITs — Failures In Time, or failures per 109 device-hours) of a large device population against age and you get a shape like a cross-section of a bathtub. It has three distinct regions, each driven by a different physical mechanism:
- Infant mortality (early life): a high but rapidly falling failure rate. Weak units with built-in defects fail first.
- Useful life (random failures): a low, roughly constant rate. Failures here are stochastic — a stray alpha particle, an ESD event, a rare oxide breakdown.
- Wear-out: a rising rate as cumulative damage mechanisms (electromigration, TDDB, NBTI, hot-carrier injection) finally consume the design margin.
Only the left region is addressed by burn-in. Burn-in does not improve a part's wear-out limit — if anything it consumes a little of that margin. Its sole job is to walk the population down the steep infant-mortality slope before shipment.
What actually causes infant mortality
Early-life failures are not random bad luck; they are latent process defects that happen to pass a one-shot test at nominal conditions but are too marginal to survive stress. Common physical culprits include:
- Gate-oxide pinholes / thin spots: a thinned dielectric passes at 0.75 V but breaks down (TDDB) within hours at 1.1 V and 125 °C.
- Metal/via voids and thin necks: a partially formed via carries rated current at room temperature but opens quickly under electromigration stress.
- Contamination and particle defects: a mobile-ion or particle-induced resistive short that is borderline at sort.
- Marginal contacts and silicide defects that raise local current density and heat.
Because these defects accelerate strongly with voltage and temperature, a short period of elevated stress compresses what would have been weeks of field life into a few hours of factory time. That is the entire principle behind burn-in.
Burn-in: accelerating the weak to failure
In burn-in, devices are loaded into high-temperature sockets (an oven-and-board system from vendors such as Aehr, Advantest or Micro Control) and run at elevated junction temperature (typically 125–150 °C) and elevated VDD (often 1.2–1.4× nominal) while the pattern engine toggles internal nodes. The acceleration is quantified by two models product/reliability engineers must know:
- Arrhenius (thermal): AF = exp[ (Ea/k) · (1/Tuse − 1/Tstress) ]. With an activation energy Ea=0.7 eV, moving from 55 °C use to 125 °C stress gives an acceleration factor of roughly 40–50×.
- Voltage (field-driven, e.g. oxide): AF = exp[ γ·(Vstress − Vuse) ], contributing another several-fold multiplier.
With a combined acceleration factor near 200×, a 12-hour burn-in reproduces roughly 100 days of field operation at use conditions — enough to push the vast majority of infant-mortality parts past their failure point so they can be rejected at post-burn-in final test.
Cost, DPPM and knowing when to stop
Burn-in is expensive — sockets, ovens, hours of tester time and yield loss all add up — so not every product gets it. The decision is a DPPM (defective parts per million) trade-off driven by the end application:
| Segment | Target field DPPM | Burn-in policy |
|---|---|---|
| Consumer (toy, low-cost MCU) | ~500–1000 | Often none; rely on test coverage |
| Mobile / compute SoC | ~50–100 | Sample or targeted burn-in |
| Automotive (AEC-Q100 Grade 0/1) | < 1 (zero-defect goal) | 100% burn-in common |
| Medical / aerospace / defence | < 0.1 | Extended burn-in + traceability |
Modern high-volume lines increasingly replace blanket burn-in with screening analytics: outlier detection methods such as Part Average Testing (PAT), Statistical Bin Limits (SBL) and IDDQ deltas flag the statistically weird dies that are the most likely infant-mortality candidates, so burn-in is applied selectively or supplemented by on-die reliability monitors and system-level test. Data flows from ATE platforms (Teradyne, Advantest) into fab-wide analytics so that a die that is numerically an outlier — even though it "passed" — can still be purged. Design-for-test and reliability simulation work in tools like Synopsys and Siemens EDA feed directly into these screens.
Where this fits in the flow
Infant mortality, random, and wear-out failures are three separate problems with three separate countermeasures. Burn-in and screening own the left of the curve; derating, redundancy and ECC handle the middle; and physics-of-failure margining (electromigration budgets, Vmax limits, mission-profile analysis) protects the right. A reliability engineer has to reason across all three — a point we expand on in our companion piece on burn-in testing methodology and in our deep dive on accelerated life testing. For hands-on PCB-level and chip-level validation flows you can also explore design-services context from Avecas VLSI design services.
Want to master the full post-silicon reliability stack — bathtub modeling, burn-in economics, DPPM screens and qualification? Our Post-Silicon Validation track on CourseTron courses walks you through real datasheets and ATE logs.
Frequently asked questions
Does burn-in reduce a device's total lifetime?
Slightly, yes. Burn-in consumes a few hours of the useful-life and wear-out budget. For a part rated for 10+ years of field life that consumption is negligible, and it is vastly outweighed by removing the infant-mortality population.
Why not just test harder instead of burning in?
Standard test is a snapshot at one instant. Many latent defects (oxide thin spots, resistive vias) are time- and stress-dependent — they simply do not manifest until the device has operated under stress for a while. Only an accelerated time-stress exposure like burn-in reveals them.
What is a typical infant-mortality duration in the field?
It varies by technology, but the falling slope of the bathtub typically flattens out within the first few hundred to a few thousand power-on hours — which is exactly why a few hundred equivalent hours of accelerated burn-in is effective.
Is burn-in still used at advanced nodes?
Blanket burn-in is increasingly replaced by data-driven outlier screening (PAT/SBL/IDDQ) plus on-chip aging sensors, but safety-critical automotive and medical parts still rely heavily on it to hit near-zero DPPM targets.