Semiconductor Manufacturing Process Measurement Uncertainty: Mathematical Modeling

Keywords: measurement uncertainty, metrology, GUM, type A uncertainty, type B uncertainty, uncertainty propagation

Semiconductor Manufacturing Process Measurement Uncertainty: Mathematical Modeling

1. The Fundamental Challenge

At modern nodes (3nm, 2nm), we face a profound problem: measurement uncertainty can consume 30–50% of the tolerance budget.

Consider typical values:

• Feature dimension: ~15nm
• Tolerance: ±1nm (≈7% variation allowed)
• Measurement repeatability: ~0.3–0.5nm
• Reproducibility (tool-to-tool): additional 0.3–0.5nm

This means we cannot naively interpret measured variation as process variation—a significant portion is measurement noise.

2. Variance Decomposition Framework

The foundational mathematical structure is the decomposition of total observed variance:

2.1 Hierarchical Decomposition

For a full fab model:

Where:

The variance components:

Measurement system variance:

3. Gauge R&R Mathematics

The standard Gauge Repeatability and Reproducibility analysis partitions measurement variance:

3.1 Key Metrics

Precision-to-Tolerance Ratio:

where $k = 5.15$ (99% coverage) or $k = 6$ (99.73% coverage)

Discrimination Ratio:

This gives the number of distinct categories the measurement system can reliably distinguish.

• Industry standard requires: $\text{ndc} \geq 5$

Signal-to-Noise Ratio:

4. GUM-Based Uncertainty Propagation

Following the Guide to the Expression of Uncertainty in Measurement (GUM):

4.1 Combined Standard Uncertainty

For a measurand $y = f(x_1, x_2, \ldots, x_n)$:

4.2 Type A vs. Type B Uncertainties

Type A (statistical):

Type B (other sources):

• Calibration certificates: $u_B = \frac{U}{k}$ where $U$ is expanded uncertainty
• Rectangular distribution (tolerance): $u_B = \frac{a}{\sqrt{3}}$
• Triangular distribution: $u_B = \frac{a}{\sqrt{6}}$

5. Spatial Modeling of Within-Wafer Variation

Within-wafer variation often has systematic spatial structure that must be separated from random measurement error.

5.1 Polynomial Surface Model (Zernike Polynomials)

Using Zernike polynomials—natural for circular wafer geometry:

• $Z_0^0$: piston (mean)
• $Z_1^1$: tilt
• $Z_2^0$: defocus (bowl shape)
• Higher orders: astigmatism, coma, spherical aberration analogs

5.2 Gaussian Process Model

For flexible, non-parametric spatial modeling:

With squared exponential covariance:

Where:

• $\sigma^2_f$: process variance (spatial signal)
• $\ell$: length scale (spatial correlation distance)
• $\sigma^2_n$: measurement noise (nugget effect)

This naturally separates spatial process variation from measurement noise.

6. Bayesian Hierarchical Modeling

Bayesian approaches provide natural uncertainty quantification and handle small samples common in expensive semiconductor metrology.

6.1 Basic Hierarchical Model

Level 1 (within-wafer measurements):

Level 2 (wafer-to-wafer variation):

Level 3 (hyperpriors):

6.2 Posterior Inference

The posterior distribution:

Solved via MCMC methods:

• Gibbs sampling
• Hamiltonian Monte Carlo (HMC)
• No-U-Turn Sampler (NUTS)

7. Monte Carlo Uncertainty Propagation

For complex, non-linear measurement models where analytical propagation fails:

7.1 Algorithm (GUM Supplement 1)

1. Define probability distributions for all input quantities $X_i$ 2. Sample $M$ realizations: $\{x_1^{(k)}, x_2^{(k)}, \ldots, x_n^{(k)}\}$ for $k = 1, \ldots, M$ 3. Propagate each sample: $y^{(k)} = f(x_1^{(k)}, \ldots, x_n^{(k)})$ 4. Analyze output distribution to obtain uncertainty

Typically $M \geq 10^6$ for reliable coverage interval estimation.

7.2 Application: OCD (Optical CD) Metrology

Scatterometry fits measured spectra to electromagnetic models with parameters:

• CD (critical dimension)
• Sidewall angle
• Height
• Layer thicknesses
• Optical constants

The measurement equation is highly non-linear:

Monte Carlo propagation captures correlations and non-linearities that linearized GUM misses.

8. The Deconvolution Problem

Given observed data that is a convolution of true process variation and measurement noise:

Goal: Recover $f_{\text{true}}$ given $f_{\text{obs}}$ and knowledge of $f_{\text{meas}}$.

8.1 Fourier Approach

In frequency domain:

Naively:

Problem: Ill-posed—small errors in $\hat{f}_{\text{obs}}$ amplified where $\hat{f}_{\text{meas}}$ is small.

8.2 Regularization Techniques

Tikhonov regularization:

Bayesian approach:

With appropriate priors (smoothness, non-negativity) to regularize the solution.

9. Virtual Metrology with Uncertainty Quantification

Virtual metrology predicts measurements from process tool data, reducing physical sampling requirements.

9.1 Model Structure

Where $\mathbf{x}_{\text{FDC}}$ = fault detection and classification data (temperatures, pressures, flows, RF power, etc.)

9.2 Uncertainty-Aware ML Approaches

Gaussian Process Regression:

Provides natural predictive uncertainty:

Conformal Prediction:

Distribution-free prediction intervals:

Where $\hat{q}$ is calibrated on held-out data to guarantee coverage probability.

10. Control Chart Implications

Measurement uncertainty affects statistical process control profoundly.

10.1 Inflated Control Limits

Standard control chart limits:

But $\sigma_{\bar{x}}$ includes measurement variance:

10.2 Adjusted Process Capability

True process capability:

Must correct observed variance:

Warning: This can yield negative estimates if measurement variance dominates—indicating the measurement system is inadequate.

11. Multi-Tool Matching and Reference Frame

11.1 Tool-to-Tool Bias Model

Where $\beta_k$ is systematic bias for tool $k$.

11.2 Mixed-Effects Formulation

• $\tau_i$: true sample value (random)
• $t_j$: tool effect (random or fixed)
• $\epsilon_{ij}$: residual

REML (Restricted Maximum Likelihood) estimation separates these components.

11.3 Traceability Chain

Total reference uncertainty:

12. Practical Uncertainty Budget Example

For CD-SEM measurement of a 20nm line:

Combined standard uncertainty:

Expanded uncertainty ($k=2$, 95% confidence):

For a ±1nm tolerance, this means P/T ≈ 60%—marginally acceptable.

13. Key Takeaways

The mathematical modeling of measurement uncertainty in semiconductor manufacturing requires:

1. Hierarchical variance decomposition (ANOVA, mixed models) to separate process from measurement variation

2. Spatial statistics (Gaussian processes, Zernike decomposition) for within-wafer systematic patterns

3. Bayesian inference for rigorous uncertainty quantification with limited samples

4. Monte Carlo methods for non-linear measurement models (OCD, model-based metrology)

5. Deconvolution techniques to recover true process distributions

6. Machine learning with uncertainty for virtual metrology

The Fundamental Insight

At nanometer scales, measurement uncertainty is not a nuisance to be ignored—it is a primary object of study that directly determines our ability to control and optimize semiconductor processes.

Key Equations Quick Reference

Variance Decomposition

GUM Combined Uncertainty

where $c_i = \frac{\partial f}{\partial x_i}$ are sensitivity coefficients.

Precision-to-Tolerance Ratio

Process Capability (Corrected)

Notation Reference

Source: ChipFoundryServices — Search this topic — Ask CFSGPT