Semiconductor Fabrication

D0 (Defect Density)

Q: How does D0 affect the yield of a large MMIC die?

Yield falls roughly exponentially with the product of D0 and die critical area. Using the Poisson model Y = exp(-A times D0), a 4 mm by 4 mm power MMIC has a die area of 0.16 cm2. At a mature GaAs D0 of 0.3 defects per cm2 the random-limited yield is exp(-0.048) = 95.3 percent. If the same process degrades to D0 = 2 per cm2, yield drops to exp(-0.32) = 72.6 percent. Because area enters the exponent, doubling die size has the same yield impact as doubling D0, which is why large millimeter-wave power dies are so sensitive to defect density and why RF foundries push hard to keep D0 low.

Q: What is the difference between the Murphy and Poisson yield models for D0?

The simple Poisson model assumes defects are perfectly randomly distributed and gives Y = exp(-A times D0). Real wafers show defect clustering, so the pure Poisson model is pessimistic for large dies. The Murphy model integrates over a distribution of defect densities and, with a triangular distribution, yields Y = ((1 - exp(-A times D0)) / (A times D0)) squared. The negative binomial model Y = (1 + A times D0 / alpha) raised to -alpha adds a clustering parameter alpha (typically 2 to 5) and is the form most fabs actually fit to data. All three share D0 as the core defect-density input; they differ only in how they treat clustering.

Q: How is D0 extracted from production wafer data?

D0 is back-calculated from measured die yield and known critical area, not measured directly. Engineers fit yield versus die area across several products on the same process, then solve the chosen yield model for D0. Inline defect inspection tools (brightfield, darkfield, and e-beam scanners) count and size particles, and only defects large enough to land in a critical area become killer defects that contribute to D0. Separating the systematic yield floor (mask, design, and parametric losses) from the random component is essential, otherwise D0 is overstated. A typical mature 0.15 um GaAs pHEMT MMIC line runs D0 in the 0.2 to 0.8 defects per cm2 range.

/dee-zee-roh/

Pronounced "D-zero," this symbol denotes the random killer-defect density of a wafer process, the average number of yield-limiting defects per square centimeter. It is the master cleanliness metric of a fab and the primary input to the Murphy and Poisson defect density yield models, which predict die yield as Y ≈ e^−A·D₀ for a die of critical area A. Only defects large enough to fall on a sensitive feature count as a killer defect and contribute to D₀; benign particles do not. A mature 0.15 μm GaAs pHEMT MMIC line typically runs D₀ of 0.2 to 0.8 per cm², while a state-of-the-art silicon CMOS node targets values below 0.1 per cm². Because die area sits in the exponent, large millimeter-wave power dies are punished severely by even modest increases in D₀, which makes defect-density reduction the single highest-impact activity in any RF foundry.

Category: Semiconductor Fabrication

Units: defects / cm²

Mature GaAs MMIC: 0.2 to 0.8 / cm²

How D0 Drives Semiconductor Die Yield

D₀ is a fab-level statistic, not a single measurable defect. It captures the density of randomly distributed, yield-killing imperfections such as particles, pinholes, micro-scratches, and shorts that land somewhere on the wafer during processing. Because these defects are spatially random, their effect on a given die follows a probability distribution rather than a fixed loss. The simplest description is the Poisson model: if a die presents critical area A to the defect population, the expected number of killer defects per die is A·D₀, and the probability that a die receives zero killer defects, which is its random-limited yield, is e^−A·D₀. This single relationship is why D₀ appears in nearly every fab cost model.

Real wafers do not obey perfect randomness. Defects cluster near the wafer edge, along scratch tracks, or in a single contaminated lot, so the pure Poisson model is pessimistic for large dies. Murphy addressed this by integrating yield over a distribution of local defect densities; with a triangular distribution the result is Y = ((1 − e^−A·D₀) / (A·D₀))². Production fabs more commonly fit the negative-binomial model, which folds clustering into a single parameter α. In all of these, D₀ is the constant that ties measured yield back to process cleanliness, so it must be separated cleanly from the systematic yield floor caused by mask, design, and parametric losses.

For RF and millimeter-wave products the stakes are higher than for digital logic. A GaAs or GaN power MMIC may occupy several square millimeters of active area packed with air bridges, via holes, and thin-film resistors, every one of which adds critical area. A defect-density excursion that a small digital die would shrug off can collapse the yield of a large power die, driving up cost per good die and stretching delivery schedules. Tracking D₀ by layer and by defect mode lets process engineers target the dominant contributor rather than chasing every particle equally.

The Yield Equations Built on D0

Poisson (random defects):
Y = e^{−A × D₀}

Murphy (triangular density distribution):
Y = ( (1 − e^{−A × D₀}) / (A × D₀) )²

Negative binomial (clustered defects):
Y = (1 + A × D₀ / α)^−α

Where A = die critical area (cm²), D₀ = killer-defect density (cm⁻²), and α = clustering parameter (typically 2 to 5; α → ∞ reduces to Poisson). Example: A = 0.16 cm², D₀ = 0.3 cm⁻² → Poisson Y ≈ e^−0.048 ≈ 95.3%.

Typical D0 Across Process Technologies

Process	Typical D₀ (cm⁻²)	Maturity	Common Killer Defect	RF Relevance
Leading-edge Si CMOS	0.05 to 0.10	Ramping to mature	Particle / bridging short	RF-SOI, RF front ends
Mature Si BiCMOS / SiGe	0.1 to 0.3	Mature	Via void, metal short	mmWave transceivers
GaAs pHEMT MMIC (0.15 μm)	0.2 to 0.8	Mature	Air-bridge collapse, via	LNAs, mixers, switches
GaN-on-SiC HEMT	0.5 to 2.0	Maturing	Crystal dislocation, pit	High-power amplifiers
Early process bring-up	2 to 10+	Development	Multiple modes	All new RF nodes

Common Questions

Frequently Asked Questions

How does D0 affect the yield of a large MMIC die?

Yield falls roughly exponentially with A × D₀. A 4 mm × 4 mm power MMIC has 0.16 cm² of die area; at a mature GaAs D₀ = 0.3 cm⁻² the Poisson yield is e^−0.048 ≈ 95.3%. Let D₀ drift to 2 cm⁻² and yield collapses to e^−0.32 ≈ 72.6%. Because area is in the exponent, doubling die size hurts as much as doubling D₀.

What is the difference between the Murphy and Poisson yield models for D0?

Poisson assumes perfectly random defects: Y = e^−A·D₀, which is pessimistic for large dies because real defects cluster. Murphy integrates over a triangular density distribution, giving Y = ((1 − e^−A·D₀) / (A·D₀))². The negative-binomial model adds a clustering parameter α and is the form most fabs fit to real data. All three use D₀ as the core input.

How is D0 extracted from production wafer data?

D₀ is back-calculated, not measured directly. Engineers fit die yield versus critical area across several products on one process, then solve the chosen yield model for D₀. Inline brightfield, darkfield, and e-beam inspection count and size particles; only defects large enough to hit a critical area count as killers. Separating systematic losses from the random component is essential or D₀ is overstated.

Related Terms

← D-Mode HEMT DAA →