Advanced Fitting Tutorial

Introduction

This advanced tutorial explores the optimization algorithms used for thickness extraction, focusing on when and how to use each method effectively. We'll cover parameter selection, convergence diagnostics, and troubleshooting common issues.

The Optimization Problem

Objective Function

All thickness extraction methods minimize the difference between measured and modeled reflectance:

χ²(d) = Σᵢ [R_meas(λᵢ) - R_model(λᵢ, d)]²

Where:

d: Film thickness (parameter to optimize)
λᵢ: Wavelength measurement points
R_meas: Measured reflectance spectrum
R_model: Theoretical reflectance from transfer matrix

Why Optimization is Needed

Unlike FFT which gives a direct estimate, finding the exact thickness requires solving a nonlinear inverse problem. The reflectance R(λ, d) has a complex functional form involving:

Trigonometric functions (cosine from interference)
Wavelength-dependent refractive indices (dispersion)
Multiple Fresnel reflections

No closed-form solution exists → iterative optimization required.

Algorithm 1: Levenberg-Marquardt (L-M)

Mathematical Foundation

L-M is a damped least-squares method that interpolates between gradient descent and Gauss-Newton approaches:

d_k+1 = d_k - [J^TJ + λI]^-1 J^Tr

Where:

J: Jacobian matrix (∂R/∂d at each wavelength)
r: Residual vector (R_meas - R_model)
λ: Damping parameter (adjusted each iteration)
I: Identity matrix

How It Works

Start: Initial guess d₀ (typically from FFT)
Compute: Calculate R_model(λ, d_k) and residuals
Gradient: Numerical Jacobian: J ≈ [R(d+h) - R(d-h)] / 2h
Update: Solve linear system for step direction
Accept/Reject: If χ² decreases, accept step and reduce λ. Else, increase λ and retry
Converge: Stop when |d_k+1 - d_k| < tolerance

Parameter Selection

Parameter	Recommended Value	Description
Initial λ	0.001-0.01	Smaller for good initial guess
λ increase	×10	When step rejected
λ decrease	÷10	When step accepted
Max iterations	50-100	Typical convergence: 10-20 iterations
Tolerance	0.01 nm	Stop criterion for parameter change

When to Use

✓ Advantages:

Very fast convergence (0.1 seconds typical)
High precision (±0.5 nm)
Provides uncertainty estimate from Jacobian
Standard method in commercial tools

✗ Limitations:

Requires good initial guess (within ~20% of true value)
Can converge to local minima
Sensitive to measurement noise at edges

Best Practice

Always use FFT estimate as initial guess for L-M. This hybrid approach combines FFT's robustness with L-M's precision.

Algorithm 2: Differential Evolution (DE)

Population-Based Global Search

DE is a stochastic evolutionary algorithm that maintains a population of candidate solutions and evolves them toward the global optimum.

Algorithm Steps

Initialize: Create population of NP random thickness values within bounds
Mutation: For each individual xᵢ, create mutant:
vᵢ = x_r1 + F(x_r2 - x_r3)
where r1, r2, r3 are random distinct indices
Crossover: Mix parent and mutant with probability CR:
uᵢⱼ = vᵢⱼ if rand() < CR, else uᵢⱼ = xᵢⱼ
Selection: Keep offspring if χ²(offspring) < χ²(parent)
Iterate: Repeat for G generations until convergence

Hyperparameters

Parameter	Typical Value	Effect
Population size (NP)	10-20	Larger = better global search, slower
Mutation factor (F)	0.5-1.0	Step size in parameter space
Crossover rate (CR)	0.7-0.9	Probability of using mutant
Max generations	100-300	Budget for convergence

When to Use

✓ Advantages:

No initial guess needed
Finds global minimum (avoids local minima)
Robust to noisy data
Simple to implement

✗ Limitations:

Slower (1-2 seconds typical)
More function evaluations (300-500 vs 10-20 for L-M)
Stochastic (slightly different result each run)
No uncertainty estimate from algorithm itself

When DE is Essential

Unknown samples: If you don't know approximate thickness (no FFT possible due to too few fringes), DE can search the entire physically reasonable range.

Multi-modal problems: Some configurations have multiple local minima. DE explores globally and avoids getting trapped.

Algorithm Comparison

Levenberg-Marquardt

Type: Local gradient-based

Speed: ⚡⚡⚡ Fast (0.1s)

Accuracy: ★★★ Excellent (±0.5 nm)

Robustness: ★★☆ Needs good guess

Best for: Production QC, high-throughput

Differential Evolution

Type: Global stochastic

Speed: ⚡☆☆ Slow (1.5s)

Accuracy: ★★★ Excellent (±0.5 nm)

Robustness: ★★★ Very robust

Best for: Unknown samples, R&D

Decision Tree

Do you have FFT estimate? ─┐
                            │
                    ┌───────┴───────┐
                    │               │
                   YES             NO
                    │               │
             Use L-M with      Use DE with
             FFT as guess      wide bounds
             (fast & accurate) (slower but robust)
                    │               │
                    └───────┬───────┘
                            │
                    Does it converge?
                            │
                    ┌───────┴───────┐
                   YES             NO
                    │               │
                  Done!         Try DE or
                            adjust bounds
            

Setting Parameter Bounds

Why Bounds Matter

Proper bounds prevent unphysical solutions and improve convergence:

Physical constraints: Thickness cannot be negative or unreasonably large
Search space reduction: Smaller range = faster convergence
Regularization: Prevents optimizer from wandering into bad regions

Recommended Bounds

Situation	Lower Bound	Upper Bound
With FFT estimate d_FFT	0.7 × d_FFT	1.3 × d_FFT
Process spec (e.g., 250±25 nm)	225 nm	275 nm
Unknown, visible fringes	50 nm	1000 nm
Unknown, few/no fringes	10 nm	100 nm

Bounds Too Tight

If true thickness falls outside bounds, optimizer will converge to boundary value. Check residuals - large systematic errors indicate this problem.

Adaptive Bounds Strategy

Run FFT → get d_FFT
Set bounds: [0.7 × d_FFT, 1.3 × d_FFT]
Run L-M fitting
If converged thickness is near boundary (within 5%), expand bounds and refit

Convergence Diagnostics

How to Know if Fitting Succeeded

1. Chi-Squared Value

χ² < 10⁻⁴ : Excellent fit
χ² < 10⁻³ : Good fit
χ² < 10⁻² : Acceptable
χ² > 10⁻² : Poor fit (check data/model)

2. Residual Analysis

Plot residuals: R_meas(λ) - R_model(λ)

Good: Random scatter around zero, no patterns
Bad: Systematic oscillations → wrong model or multi-layer
Bad: Linear trend → baseline drift in data

3. Convergence History

Plot χ² vs iteration number:

Good: Smooth monotonic decrease to plateau
Bad: Oscillations without decrease → diverging
Bad: Flat from start → stuck at initial guess

4. Parameter Uncertainty

From L-M covariance matrix: σ_d < 2 nm typically

Large uncertainty (>5 nm) indicates poor sensitivity or noisy data
Check SNR of spectral data
Verify sufficient number of fringes (≥3)

Troubleshooting Common Issues

Issue 1: Non-Convergence

Symptoms: Maximum iterations reached, χ² still large

Causes & Solutions:

Bad initial guess: Switch to DE or try different guess
Wrong material index: Verify refractive index value
Multi-layer film: Single-layer model invalid
Poor data quality: Check for measurement artifacts

Issue 2: Boundary Convergence

Symptoms: Fitted thickness exactly at lower or upper bound

Solutions:

Expand bounds (e.g., double the range)
Check if thickness truly outside initial guess range
Verify FFT estimate is reasonable

Issue 3: Multiple Solutions

Symptoms: Different initial guesses converge to different values

Cause: Thick films can have ambiguous fringe counting

Solutions:

Use process knowledge to eliminate unphysical solutions
Run DE to search globally
Measure at different wavelength range to resolve ambiguity

Issue 4: Slow Convergence

Symptoms: Takes >100 iterations for L-M

Solutions:

Improve initial guess (use FFT)
Reduce damping parameter λ
Check for numerical precision issues

Advanced Techniques

1. Hybrid Two-Stage Fitting

Stage 1: Global Search
  - Run DE with wide bounds [50, 1000] nm
  - Find approximate global minimum: d_DE ≈ 245 nm
  - Time: 1.5 seconds

Stage 2: Local Refinement
  - Run L-M with initial guess d_DE
  - Narrow bounds: [0.9×d_DE, 1.1×d_DE]
  - Precise result: d = 250.15 ± 0.82 nm
  - Time: 0.1 seconds

Total: 1.6 seconds for guaranteed global optimum
            

2. Multi-Start L-M

Run L-M from 5-10 different random initial guesses, keep best result:

guesses = [100, 200, 300, 400, 500]  # nm
results = []

for guess in guesses:
    d_fit, chi2 = run_LM(guess)
    results.append((d_fit, chi2))

best_d = min(results, key=lambda x: x[1])[0]
# Best result has lowest χ²
            

3. Uncertainty Quantification via Bootstrap

Estimate uncertainty by resampling data with noise:

n_bootstrap = 100
thicknesses = []

for i in range(n_bootstrap):
    # Add random noise to measured spectrum
    R_perturbed = R_measured + noise_level * randn()
    d_fit = run_fitting(R_perturbed)
    thicknesses.append(d_fit)

d_mean = mean(thicknesses)
d_std = std(thicknesses)
# Report: d = d_mean ± d_std nm
            

Summary & Recommendations

Use Case	Recommended Strategy
Production QC (known process)	FFT + L-M with tight bounds
R&D (unknown samples)	DE with wide bounds, or Hybrid
Thin films (<50 nm, few fringes)	DE with [10, 100] nm bounds
Thick films (>500 nm, many fringes)	Multi-start L-M to handle ambiguity
Critical measurements	Hybrid + Bootstrap for uncertainty

Key Takeaways

L-M is fast but needs good initial guess → always use FFT first
DE is robust for unknown samples, worth the extra time
Bounds matter: ±20-30% around estimate is ideal
Check residuals: Random = good, patterns = problems
χ² < 10⁻⁴: Target for high-quality fits

Introduction

The Optimization Problem

Objective Function

Why Optimization is Needed

Algorithm 1: Levenberg-Marquardt (L-M)

Mathematical Foundation

How It Works

Parameter Selection

When to Use

✓ Advantages:

✗ Limitations:

Best Practice

Algorithm 2: Differential Evolution (DE)

Population-Based Global Search

Algorithm Steps

Hyperparameters

When to Use

✓ Advantages:

✗ Limitations:

When DE is Essential

Algorithm Comparison

Levenberg-Marquardt

Differential Evolution

Decision Tree

Setting Parameter Bounds

Why Bounds Matter

Recommended Bounds

Bounds Too Tight

Adaptive Bounds Strategy

Convergence Diagnostics

How to Know if Fitting Succeeded

1. Chi-Squared Value

2. Residual Analysis

3. Convergence History

4. Parameter Uncertainty

Troubleshooting Common Issues

Issue 1: Non-Convergence

Issue 2: Boundary Convergence

Issue 3: Multiple Solutions

Issue 4: Slow Convergence

Advanced Techniques

1. Hybrid Two-Stage Fitting

2. Multi-Start L-M

3. Uncertainty Quantification via Bootstrap

Summary & Recommendations

Key Takeaways

Further Reading