error propagation
**Semiconductor Manufacturing Error Propagation Mathematics**
**1. Fundamental Error Propagation Theory**
For a function $f(x_1, x_2, \ldots, x_n)$ where each variable $x_i$ has uncertainty $\sigma_i$, the propagated uncertainty follows:
$$
\sigma_f^2 = \sum_{i=1}^{n} \left( \frac{\partial f}{\partial x_i} \right)^2 \sigma_i^2 + 2 \sum_{i < j} \frac{\partial f}{\partial x_i} \frac{\partial f}{\partial x_j} \, \text{cov}(x_i, x_j)
$$
For **uncorrelated errors**, this simplifies to the **Root-Sum-of-Squares (RSS)** formula:
$$
\sigma_f = \sqrt{\sum_{i=1}^{n} \left( \frac{\partial f}{\partial x_i} \right)^2 \sigma_i^2}
$$
**Applications in Semiconductor Manufacturing**
- **Critical Dimension (CD) variations**: Feature size deviations from target
- **Overlay errors**: Misalignment between lithography layers
- **Film thickness variations**: Deposition uniformity issues
- **Doping concentration variations**: Implant dose and energy fluctuations
**2. Process Chain Error Accumulation**
Semiconductor manufacturing involves hundreds of sequential process steps. Errors propagate through the chain in different modes:
**2.1 Additive Error Accumulation**
Used for overlay alignment between layers:
$$
E_{\text{total}} = \sum_{i=1}^{n} \varepsilon_i
$$
$$
\sigma_{\text{total}}^2 = \sum_{i=1}^{n} \sigma_i^2 \quad \text{(if uncorrelated)}
$$
**2.2 Multiplicative Error Accumulation**
Used for etch selectivity, deposition rates, and gain factors:
$$
G_{\text{total}} = \prod_{i=1}^{n} G_i
$$
$$
\frac{\sigma_G}{G} \approx \sqrt{\sum_{i=1}^{n} \left( \frac{\sigma_{G_i}}{G_i} \right)^2}
$$
**2.3 Error Accumulation Modes**
- **Additive**: Errors sum directly (overlay, thickness)
- **Multiplicative**: Errors compound through products (gain, selectivity)
- **Compensating**: Rare cases where errors cancel
- **Nonlinear interactions**: Complex dependencies requiring simulation
**3. Hierarchical Variance Decomposition**
Total variation decomposes across spatial and temporal hierarchies:
$$
\sigma_{\text{total}}^2 = \sigma_{\text{lot}}^2 + \sigma_{\text{wafer}}^2 + \sigma_{\text{die}}^2 + \sigma_{\text{within-die}}^2
$$
**Variance Sources by Level**
| Level | Sources |
|-------|---------|
| **Lot-to-lot** | Incoming material, chamber conditioning, recipe drift |
| **Wafer-to-wafer** | Slot position, thermal gradients, handling |
| **Die-to-die** | Across-wafer uniformity, lens field distortion |
| **Within-die** | Pattern density, microloading, proximity effects |
**Variance Component Analysis**
For $N$ measurements $y_{ijk}$ (lot $i$, wafer $j$, site $k$):
$$
y_{ijk} = \mu + L_i + W_{ij} + \varepsilon_{ijk}
$$
Where:
- $\mu$ = grand mean
- $L_i \sim N(0, \sigma_L^2)$ = lot effect
- $W_{ij} \sim N(0, \sigma_W^2)$ = wafer effect
- $\varepsilon_{ijk} \sim N(0, \sigma_\varepsilon^2)$ = residual
**4. Yield Mathematics**
**4.1 Poisson Defect Model (Random Defects)**
$$
Y = e^{-D_0 A}
$$
Where:
- $D_0$ = defect density (defects/cm²)
- $A$ = die area (cm²)
**4.2 Negative Binomial Model (Clustered Defects)**
More realistic for actual manufacturing:
$$
Y = \left( 1 + \frac{D_0 A}{\alpha} \right)^{-\alpha}
$$
Where:
- $\alpha$ = clustering parameter
- $\alpha \to \infty$ recovers Poisson model
- Smaller $\alpha$ = more clustering
**4.3 Total Yield**
$$
Y_{\text{total}} = Y_{\text{defect}} \times Y_{\text{parametric}}
$$
**4.4 Parametric Yield**
Integration over the multi-dimensional acceptable parameter space:
$$
Y_{\text{parametric}} = \int \int \cdots \int_{\text{spec}} f(p_1, p_2, \ldots, p_n) \, dp_1 \, dp_2 \cdots dp_n
$$
For Gaussian parameters with specs at $\pm k\sigma$:
$$
Y_{\text{parametric}} \approx \left[ \text{erf}\left( \frac{k}{\sqrt{2}} \right) \right]^n
$$
**5. Edge Placement Error (EPE)**
Critical metric at advanced nodes combining multiple error sources:
$$
EPE^2 = \left( \frac{\Delta CD}{2} \right)^2 + OVL^2 + \left( \frac{LER}{2} \right)^2
$$
**EPE Components**
- $\Delta CD$ = Critical dimension error
- $OVL$ = Overlay error
- $LER$ = Line edge roughness
**Extended EPE Model**
Including additional terms:
$$
EPE^2 = \left( \frac{\Delta CD}{2} \right)^2 + OVL^2 + \left( \frac{LER}{2} \right)^2 + \sigma_{\text{mask}}^2 + \sigma_{\text{etch}}^2
$$
**6. Overlay Error Modeling**
Overlay at any point $(x, y)$ is modeled as:
$$
OVL(x, y) = \vec{T} + R\theta + M \cdot \vec{r} + \text{HOT}
$$
**Overlay Components**
- $\vec{T} = (T_x, T_y)$ = Translation
- $R\theta$ = Rotation
- $M$ = Magnification
- $\text{HOT}$ = Higher-Order Terms (lens distortions, wafer non-flatness)
**Overlay Budget (RSS)**
$$
OVL_{\text{budget}}^2 = OVL_{\text{tool}}^2 + OVL_{\text{process}}^2 + OVL_{\text{wafer}}^2 + OVL_{\text{mask}}^2
$$
**10-Parameter Overlay Model**
$$
\begin{aligned}
dx &= T_x + R_x \cdot y + M_x \cdot x + N_x \cdot x \cdot y + \ldots \\
dy &= T_y + R_y \cdot x + M_y \cdot y + N_y \cdot x \cdot y + \ldots
\end{aligned}
$$
**7. Stochastic Effects in EUV Lithography**
At EUV wavelengths (13.5 nm), photon shot noise becomes fundamental.
**Photon Statistics**
Photons per pixel follow Poisson distribution:
$$
N \sim \text{Poisson}(\bar{N})
$$
$$
\sigma_N = \sqrt{\bar{N}}
$$
**Relative Dose Fluctuation**
$$
\frac{\sigma_N}{\bar{N}} = \frac{1}{\sqrt{\bar{N}}}
$$
**Stochastic Failure Probability**
$$
P_{\text{fail}} \propto \exp\left( -\frac{E}{E_{\text{threshold}}} \right)
$$
**RLS Triangle Trade-off**
- **R**esolution
- **L**ine edge roughness (LER)
- **S**ensitivity (dose)
$$
LER \propto \frac{1}{\sqrt{\text{Dose}}} \propto \frac{1}{\sqrt{N_{\text{photons}}}}
$$
**8. Spatial Correlation Modeling**
Errors are spatially correlated. Modeled using variograms or correlation functions.
**Variogram**
$$
\gamma(h) = \frac{1}{2} E\left[ (Z(x+h) - Z(x))^2 \right]
$$
**Correlation Function**
$$
\rho(h) = \frac{\text{cov}(Z(x+h), Z(x))}{\text{var}(Z(x))}
$$
**Common Correlation Models**
| Model | Formula |
|-------|---------|
| **Exponential** | $\rho(h) = \exp\left( -\frac{h}{\lambda} \right)$ |
| **Gaussian** | $\rho(h) = \exp\left( -\left( \frac{h}{\lambda} \right)^2 \right)$ |
| **Spherical** | $\rho(h) = 1 - \frac{3h}{2\lambda} + \frac{h^3}{2\lambda^3}$ for $h \leq \lambda$ |
**Implications**
- Nearby devices are more correlated → better matching for analog
- Correlation length $\lambda$ determines effective samples per die
- Extreme values are less severe than independent variation suggests
**9. Process Capability and Tail Statistics**
**Process Capability Index**
$$
C_{pk} = \min \left[ \frac{USL - \mu}{3\sigma}, \frac{\mu - LSL}{3\sigma} \right]
$$
**Defect Rates vs. Cpk (Gaussian)**
| $C_{pk}$ | PPM Outside Spec | Sigma Level |
|----------|------------------|-------------|
| 1.00 | ~2,700 | 3σ |
| 1.33 | ~63 | 4σ |
| 1.67 | ~0.6 | 5σ |
| 2.00 | ~0.002 | 6σ |
**Extreme Value Statistics**
For $n$ independent samples from distribution $F(x)$, the maximum follows:
$$
P(M_n \leq x) = [F(x)]^n
$$
For large $n$, converges to Generalized Extreme Value (GEV):
$$
G(x) = \exp\left\{ -\left[ 1 + \xi \left( \frac{x - \mu}{\sigma} \right) \right]^{-1/\xi} \right\}
$$
**Critical Insight**
For a chip with $10^{10}$ transistors:
$$
P_{\text{chip fail}} = 1 - (1 - P_{\text{transistor fail}})^{10^{10}} \approx 10^{10} \cdot P_{\text{transistor fail}}
$$
Even $P_{\text{transistor fail}} = 10^{-11}$ matters!
**10. Sensitivity Analysis and Error Attribution**
**Sensitivity Coefficient**
$$
S_i = \frac{\partial Y}{\partial \sigma_i} \times \frac{\sigma_i}{Y}
$$
**Variance Contribution**
$$
\text{Contribution}_i = \frac{\left( \frac{\partial f}{\partial x_i} \right)^2 \sigma_i^2}{\sigma_f^2} \times 100\%
$$
**Bayesian Root Cause Attribution**
$$
P(\text{cause} \mid \text{observation}) = \frac{P(\text{observation} \mid \text{cause}) \cdot P(\text{cause})}{P(\text{observation})}
$$
**Pareto Analysis Steps**
1. Compute variance contribution from each source
2. Rank sources by contribution
3. Focus improvement on top contributors
4. Verify improvement with updated measurements
**11. Monte Carlo Simulation Methods**
Due to complexity and nonlinearity, Monte Carlo methods are essential.
**Algorithm**
```
FOR i = 1 to N_samples:
1. Sample process parameters: p_i ~ distributions
2. Simulate device/circuit: y_i = f(p_i)
3. Store result: Y[i] = y_i
END FOR
Compute statistics from Y[]
```
**Key Advantages**
- Captures non-Gaussian behavior
- Handles nonlinear transfer functions
- Reveals correlations between outputs
- Provides full distribution, not just moments
**Sample Size Requirements**
For estimating probability $p$ of rare events:
$$
N \geq \frac{1 - p}{p \cdot \varepsilon^2}
$$
Where $\varepsilon$ is the desired relative error.
For $p = 10^{-6}$ with 10% error: $N \approx 10^8$ samples
**12. Design-Technology Co-Optimization (DTCO)**
Error propagation feeds back into design rules:
$$
\text{Design Margin} = k \times \sigma_{\text{total}}
$$
Where $k$ depends on required yield and number of instances.
**Margin Calculation**
For yield $Y$ over $N$ instances:
$$
k = \Phi^{-1}\left( Y^{1/N} \right)
$$
Where $\Phi^{-1}$ is the inverse normal CDF.
**Example**
- Target yield: 99%
- Number of gates: $10^9$
- Required: $k \approx 7\sigma$ per gate
**13. Key Mathematical Insights**
**Insight 1: RSS Dominates Budgets**
Uncorrelated errors add in quadrature:
$$
\sigma_{\text{total}} = \sqrt{\sigma_1^2 + \sigma_2^2 + \cdots + \sigma_n^2}
$$
**Implication**: Reducing the largest contributor gives the most improvement.
**Insight 2: Tails Matter More Than Means**
High-volume manufacturing lives in the $6\sigma$ tails where:
- Gaussian assumptions break down
- Extreme value statistics become essential
- Rare events dominate yield loss
**Insight 3: Nonlinearity Creates Surprises**
Even Gaussian inputs produce non-Gaussian outputs:
$$
Y = f(X) \quad \text{where } X \sim N(\mu, \sigma^2)
$$
If $f$ is nonlinear, $Y$ is not Gaussian.
**Insight 4: Correlations Can Help or Hurt**
- **Positive correlations**: Worsen tail probabilities
- **Negative correlations**: Can provide compensation
- **Designed-in correlations**: Can dramatically improve yield
**Insight 5: Scaling Amplifies Relative Error**
$$
\text{Relative Error} = \frac{\sigma}{\text{Feature Size}}
$$
A 1 nm variation:
- 5% of 20 nm feature
- 10% of 10 nm feature
- 20% of 5 nm feature
**14. Summary Equations**
**Core Error Propagation**
$$
\sigma_f^2 = \sum_i \left( \frac{\partial f}{\partial x_i} \right)^2 \sigma_i^2
$$
**Yield (Negative Binomial)**
$$
Y = \left( 1 + \frac{D_0 A}{\alpha} \right)^{-\alpha}
$$
**Edge Placement Error**
$$
EPE = \sqrt{\left( \frac{\Delta CD}{2} \right)^2 + OVL^2 + \left( \frac{LER}{2} \right)^2}
$$
**Process Capability**
$$
C_{pk} = \min \left[ \frac{USL - \mu}{3\sigma}, \frac{\mu - LSL}{3\sigma} \right]
$$
**Stochastic LER**
$$
LER \propto \frac{1}{\sqrt{N_{\text{photons}}}}
$$