Semiconductor Manufacturing Process Metrology: Science, Mathematics, and Modeling

Keywords: metrology, semiconductor metrology, measurement, characterization, ellipsometry, scatterometry

Semiconductor Manufacturing Process Metrology: Science, Mathematics, and Modeling

A comprehensive exploration of the physics, mathematics, and computational methods underlying nanoscale measurement in semiconductor fabrication.

1. The Fundamental Challenge

Modern semiconductor manufacturing produces structures with critical dimensions of just a few nanometers. At leading-edge nodes (3nm, 2nm), we are measuring features only 10–20 atoms wide.

Key Requirements

• Sub-angstrom precision in measurement
• Complex 3D architectures: FinFETs, Gate-All-Around (GAA) transistors, 3D NAND (200+ layers)
• High throughput: seconds per measurement in production
• Multi-parameter extraction: distinguish dozens of correlated parameters

Metrology Techniques Overview

2. Physics Foundation

2.1 Maxwell's Equations

At the heart of optical metrology lies the solution to Maxwell's equations:

Where:

• $\mathbf{E}$ = Electric field vector
• $\mathbf{H}$ = Magnetic field vector
• $\mathbf{D}$ = Electric displacement field
• $\mathbf{B}$ = Magnetic flux density
• $\mathbf{J}$ = Current density
• $\rho$ = Charge density

2.2 Constitutive Relations

For linear, isotropic media:

The complex dielectric function:

Where:

• $n$ = Refractive index
• $k$ = Extinction coefficient

2.3 Fresnel Equations

At an interface between media with refractive indices $\tilde{n}_1$ and $\tilde{n}_2$:

s-polarization (TE):

p-polarization (TM):

With Snell's law:

3. Mathematics of Inverse Problems

3.1 Problem Formulation

Metrology is fundamentally an inverse problem:

We seek parameters $\mathbf{p}$ that minimize the difference between model $M(\mathbf{p})$ and data $\mathbf{D}$:

Or with weighted least squares:

3.2 Levenberg-Marquardt Algorithm

The workhorse optimization algorithm interpolates between gradient descent and Gauss-Newton:

Where:

• $\mathbf{J}$ = Jacobian matrix (sensitivity matrix)
• $\lambda$ = Damping parameter
• $\delta\mathbf{p}$ = Parameter update step

The Jacobian elements:

Algorithm behavior:

• Large $\lambda$ → Gradient descent (robust, slow)
• Small $\lambda$ → Gauss-Newton (fast near minimum)

3.3 Regularization Techniques

For ill-posed problems, regularization is essential:

Tikhonov Regularization (L2):

LASSO Regularization (L1):

Bayesian Inference:

Where:

• $P(\mathbf{p} | \mathbf{D})$ = Posterior probability
• $P(\mathbf{D} | \mathbf{p})$ = Likelihood
• $P(\mathbf{p})$ = Prior probability

4. Thin Film Optics

4.1 Ellipsometry Fundamentals

Ellipsometry measures the change in polarization state upon reflection:

Where:

• $\Psi$ = Amplitude ratio angle
• $\Delta$ = Phase difference
• $r_p, r_s$ = Complex reflection coefficients

4.2 Transfer Matrix Method

For multilayer stacks, the characteristic matrix for layer $j$:

Where the phase thickness:

And the optical admittance:

Total system matrix:

Reflection coefficient:

4.3 Dispersion Models

Lorentz Oscillator Model:

Tauc-Lorentz Model (for amorphous semiconductors):

With $\varepsilon_1$ obtained via Kramers-Kronig relations:

5. Scatterometry and RCWA

5.1 Rigorous Coupled-Wave Analysis

For a grating with period $\Lambda$, electromagnetic fields are expanded in Fourier orders:

Where the diffracted wave vectors:

5.2 Eigenvalue Problem

In each layer, the field satisfies:

Where $\mathbf{\Omega}^2$ is a matrix determined by the Fourier components of the permittivity:

The eigenvalue decomposition:

Provides propagation constants (eigenvalues $\lambda_m$) and field profiles (eigenvectors in $\mathbf{W}$).

5.3 S-Matrix Formulation

For numerical stability, use the scattering matrix formulation:

Where $\mathbf{a}^+$ and $\mathbf{a}^-$ represent forward and backward propagating waves.

The S-matrix is built recursively:

Using the Redheffer star product $\star$.

6. Statistical Process Control

6.1 Control Charts

$\bar{X}$ Chart (Mean):

R Chart (Range):

EWMA (Exponentially Weighted Moving Average):

With control limits:

6.2 Process Capability Indices

$C_p$ (Process Capability):

$C_{pk}$ (Centered Process Capability):

$C_{pm}$ (Taguchi Capability):

Where:

• $USL$ = Upper Specification Limit
• $LSL$ = Lower Specification Limit
• $T$ = Target value
• $\mu$ = Process mean
• $\sigma$ = Process standard deviation

6.3 Gauge R&R Analysis

Total measurement variance decomposition:

Precision-to-Tolerance Ratio:

7. Uncertainty Quantification

7.1 Fisher Information Matrix

The Fisher Information Matrix for parameter estimation:

Or equivalently:

Where $L$ is the likelihood function.

7.2 Cramér-Rao Lower Bound

The covariance matrix of any unbiased estimator is bounded:

For a single parameter:

Interpretation:

• Diagonal elements of $\mathbf{F}^{-1}$ give minimum variance for each parameter
• Off-diagonal elements indicate parameter correlations
• Large condition number of $\mathbf{F}$ indicates ill-conditioning

7.3 Correlation Coefficient

7.4 GUM Framework

According to the Guide to the Expression of Uncertainty in Measurement:

Combined standard uncertainty:

Expanded uncertainty:

Where $k$ is the coverage factor (typically $k=2$ for 95% confidence).

8. Machine Learning in Metrology

8.1 Neural Network Surrogate Models

Replace expensive physics simulations with trained neural networks:

Training objective:

Speedup: Typically $10^4$ – $10^6 \times$ faster than RCWA/FEM.

8.2 Physics-Informed Neural Networks (PINNs)

Incorporate physical laws into the loss function:

Where:

8.3 Gaussian Process Regression

A non-parametric Bayesian approach:

Common kernel (RBF/Squared Exponential):

Posterior prediction:

Advantages:

• Provides uncertainty estimates naturally
• Works well with limited training data
• Interpretable hyperparameters

8.4 Virtual Metrology

Predict wafer properties from equipment sensor data:

Where $FDC_i$ are Fault Detection and Classification sensor readings.

Common approaches:

• Partial Least Squares (PLS) regression
• Random Forests
• Gradient Boosting (XGBoost, LightGBM)
• Deep neural networks

9. Advanced Topics and Frontiers

9.1 3D Metrology Challenges

Modern structures require 3D measurement:

9.2 Hybrid Metrology

Combining multiple techniques to break parameter correlations:

Example combination:

• OCD for periodic structure parameters
• Ellipsometry for film optical constants
• XRR for density and interface roughness

Mathematical framework:

Reduces off-diagonal elements, improving condition number.

9.3 Atomic-Scale Considerations

At the 2nm node and beyond:

Line Edge Roughness (LER):

Power Spectral Density:

Where:

• $\xi$ = Correlation length
• $H$ = Hurst exponent (roughness character)

Quantum Effects:

• Tunneling through thin barriers
• Discrete dopant effects
• Wave function penetration

9.4 Model-Measurement Circularity

A fundamental epistemological challenge:

-
┌──────────────┐      ┌──────────────┐
│   Physical   │ ───► │   Measured   │
│   Structure  │      │    Signal    │
└──────────────┘      └──────────────┘
       ▲                     │
       │                     ▼
       │              ┌──────────────┐
       │              │    Model     │
       └────────────◄─┤   Inversion  │
                      └──────────────┘

Key questions:

• How do we validate models when "truth" requires modeling?
• Reference metrology (TEM) also requires interpretation
• What does it mean to "know" a dimension at atomic scale?

Key Symbols and Notation

Computational Complexity

Where:

• $N$ = Number of layers / mesh elements
• $M$ = Number of Fourier orders
• $L$ = Number of layers
• $T$ = Number of time steps

Source: ChipFoundryServices — Search this topic — Ask CFSGPT