Negative Binomial Yield Model

Keywords: negative binomial yield model,manufacturing

Negative Binomial Yield Model is the industry-standard yield prediction framework that accounts for spatial clustering of defects — extending the Poisson model with a clustering parameter α that captures the non-random, clustered distribution of real manufacturing defects, providing significantly more accurate yield estimates — the model used by every major semiconductor fab for production yield prediction, capacity planning, and die cost estimation because it matches empirical yield data far better than the random-defect Poisson assumption.

What Is the Negative Binomial Yield Model?

- Definition: Y = [1 + (D₀ × A) / α]⁻α, where Y is die yield, D₀ is average defect density, A is die area, and α is the clustering parameter that describes how spatially clustered defects are on the wafer.
- Clustering Parameter α: Controls the degree of defect spatial correlation — α → ∞ recovers the Poisson model (random defects), α → 0 represents severe clustering where defects concentrate in patches.
- Physical Interpretation: In a wafer with clustered defects, some regions are heavily contaminated while other regions are nearly defect-free — this clustering actually improves yield compared to the random (Poisson) case because more die escape defect-heavy zones entirely.
- Typical α Values: α = 0.5–2.0 for mature processes; α = 0.3–0.5 for immature or defect-prone processes; α > 5 approaches Poisson behavior.

Why the Negative Binomial Model Matters

- Accurate Yield Prediction: Matches empirical yield data within 1–3% absolute for mature fabs — the Poisson model can be off by 10–20% for large die due to ignoring clustering.
- Revenue Forecasting: Accurate yield prediction feeds die-per-wafer output calculations that determine fab revenue — a 5% yield prediction error on high-volume products means millions in forecasting error.
- Capacity Planning: Wafer starts required = demand / (dies per wafer × yield) — accurate yield models prevent both over-investment and under-delivery.
- Process Maturity Tracking: The α parameter tracks process maturity independently of D₀ — improving α indicates better defect spatial uniformity even if total defect density hasn't changed.
- Die Size Optimization: The negative binomial model more accurately captures the area-yield relationship — critical for reticle layout decisions balancing die size against yield.

Negative Binomial vs. Poisson Comparison

| D₀ × A | Poisson Yield | NB Yield (α=0.5) | NB Yield (α=2.0) |
|---------|--------------|-------------------|-------------------|
| 0.1 | 90.5% | 90.9% | 90.7% |
| 0.5 | 60.7% | 66.7% | 63.0% |
| 1.0 | 36.8% | 50.0% | 42.0% |
| 2.0 | 13.5% | 33.3% | 23.6% |
| 5.0 | 0.7% | 14.3% | 6.3% |

Key Insight: Clustering (lower α) actually improves yield compared to random defects — because defects pile up in "bad zones" leaving more die in "good zones" completely defect-free.

Extracting Model Parameters

From Wafer Sort Data:
- Measure die pass/fail across multiple wafers.
- Fit yield vs. die-area data to negative binomial model using maximum likelihood estimation.
- Extract D₀ (average defect density) and α (clustering parameter) simultaneously.

From Defect Inspection:
- Map defect coordinates from inspection tools (KLA, Applied Materials).
- Calculate spatial clustering statistics (Moran's I, nearest-neighbor index).
- Convert clustering metrics to equivalent α parameter.

Process Maturity Stages

| Development Phase | Typical D₀ | Typical α | Yield (1 cm² die) |
|-------------------|-----------|-----------|-------------------|
| Early Development | >5 /cm² | 0.3–0.5 | <15% |
| Process Qualification | 1–2 /cm² | 0.5–1.0 | 30–50% |
| Volume Ramp | 0.3–1.0 /cm² | 1.0–2.0 | 50–75% |
| Mature Production | <0.3 /cm² | 1.5–3.0 | >80% |

Negative Binomial Yield Model is the quantitative backbone of semiconductor manufacturing economics — providing the accurate yield predictions that drive wafer start decisions, capacity investments, product pricing, and profitability analysis, making it the most important equation in the business of semiconductor fabrication.

Want to learn more?

Search 13,225+ semiconductor and AI topics or chat with our AI assistant.

Search Topics Chat with CFSGPT