Lithography Yield & Quality

Comprehensive guide to DOE, SPC, process capability, and continuous improvement methods

1. Design of Experiments (DOE)

Full Factorial Designs

Full factorial DOE systematically varies all factors at all levels to explore the complete experimental space. For k factors each at 2 levels, a full factorial requires 2^k runs. This enables estimation of all main effects and interactions.

2^k Full Factorial Example (k=3 factors):
Factors: Dose (D), Focus (F), PEB Temperature (T)
Levels: Each at 2 levels (low -, high +)
Total runs: 2³ = 8 experiments

Run Design Matrix:
Run 1: D-, F-, T- | Run 5: D-, F-, T+
Run 2: D+, F-, T- | Run 6: D+, F-, T+
Run 3: D-, F+, T- | Run 7: D-, F+, T+
Run 4: D+, F+, T- | Run 8: D+, F+, T+

Model: Y = β₀ + β₁D + β₂F + β₃T + β₁₂DF + β₁₃DT + β₂₃FT + β₁₂₃DFT
All main effects, 2-way interactions, and 3-way interaction estimated

Fractional Factorial Designs

When full factorial is impractical (e.g., 2^7 = 128 runs for 7 factors), fractional factorial designs reduce experiment count by confounding higher-order interactions:

Example: 2^(7-4) Fractional Factorial (Resolution III)
7 factors, 2⁷⁻⁴ = 8 runs (instead of 128)
Reduction: 16× fewer experiments

Trade-off: Main effects confounded with 2-way interactions
Use domain knowledge to assume negligible high-order interactions

Response Surface Methodology (RSM)

RSM fits quadratic models to optimize process parameters and map process windows:

Second-order model:
Y = β₀ + Σβᵢxᵢ + Σβᵢᵢxᵢ² + ΣΣβᵢⱼxᵢxⱼ + ε

Central Composite Design (CCD):
- Factorial points: 2^k corners of design space
- Axial points: 2k points at ±α along each axis
- Center points: n_c replicates for error estimation
Total runs: 2^k + 2k + n_c

Example (k=2): 4 corners + 4 axial + 5 center = 13 runs
Analysis: Fit quadratic model, optimize via gradient or contour plots

Taguchi Methods and Robust Design

Taguchi DOE focuses on parameter design to minimize sensitivity to noise factors. Uses orthogonal arrays for efficient multi-factor studies:

DOE Best Practices: (1) Randomize run order to avoid confounding with time trends, (2) Include center point replicates to estimate pure error, (3) Block experiments if runs cannot be completed in one batch, (4) Use domain knowledge to select appropriate factor ranges and identify negligible interactions.

2. Statistical Process Control (SPC)

Control Charts: X-bar and R Charts

X-bar and R charts monitor process mean and variability using subgroup data. Statistics are plotted over time with control limits at ±3σ:

X-bar Chart (monitors process mean):
X̄ᵢ = subgroup mean = (1/n) Σxᵢⱼ
Grand mean: X̿ = (1/m) ΣX̄ᵢ
Control limits: UCL = X̿ + A₂·R̄, CL = X̿, LCL = X̿ - A₂·R̄

R Chart (monitors process variability):
Rᵢ = max(subgroup i) - min(subgroup i)
R̄ = average range
Control limits: UCL = D₄·R̄, CL = R̄, LCL = D₃·R̄

Constants A₂, D₃, D₄ depend on subgroup size n (from statistical tables)

Western Electric Rules

Eight patterns detect non-random variation indicating loss of control:

  1. Rule 1: Any point beyond 3σ limits → out of control
  2. Rule 2: 2 out of 3 consecutive points beyond 2σ (same side)
  3. Rule 3: 4 out of 5 consecutive points beyond 1σ (same side)
  4. Rule 4: 8 consecutive points on same side of centerline
  5. Rule 5: 6 consecutive points monotonically increasing or decreasing
  6. Rule 6: 15 consecutive points within ±1σ
  7. Rule 7: 14 consecutive points alternating up and down
  8. Rule 8: 8 consecutive points beyond ±1σ on both sides

Process Capability Indices

Capability indices compare process variation to specification tolerance:

Cp = (USL - LSL) / (6σ) [potential capability, centering ignored]
Cpk = min[(USL - μ)/(3σ), (μ - LSL)/(3σ)] [actual capability]

Interpretation:
Cpk < 1.0: Incapable, defects likely (0.27% out-of-spec at ±3σ)
Cpk = 1.33: Capable (63 ppm defects)
Cpk = 1.67: Highly capable (0.6 ppm defects, 5σ level)
Cpk ≥ 2.0: Excellent (0.002 ppm defects, 6σ level)

Advanced lithography target: Cpk > 1.67 for critical parameters
SPC Implementation: Collect 25-30 subgroups for baseline control limits. Monitor in real-time, investigating violations immediately. Recalculate limits monthly or after verified process changes. Use SPC for continuous improvement, not just defect detection.

3. Process Capability Analysis

Cp, Cpk, Pp, Ppk Calculations

Short-term (Cp, Cpk) and long-term (Pp, Ppk) metrics provide complementary performance views:

Metric Formula Variation Source Use Case
Cp (USL - LSL) / (6σ_within) Within-subgroup Short-term potential
Cpk min[(USL - μ)/(3σ_within), (μ - LSL)/(3σ_within)] Within-subgroup Short-term actual
Pp (USL - LSL) / (6σ_overall) Overall (all data) Long-term potential
Ppk min[(USL - μ)/(3σ_overall), (μ - LSL)/(3σ_overall)] Overall (all data) Long-term actual

Six Sigma Methodology and DPMO

Six Sigma targets Cpk ≥ 2.0 (process mean at least 6σ from nearest spec limit):

Sigma Level Cpk DPMO (1.5σ shift) Yield
1.00 66,807 93.3%
1.33 6,210 99.38%
1.67 233 99.977%
2.00 3.4 99.99966%
DPMO = (Number of Defects / (Units × Opportunities per Unit)) × 1,000,000

Example: 25 wafers × 500 die × 20 critical layers = 250,000 opportunities
If 2,500 defective die: DPMO = (2,500 / 250,000) × 1M = 10,000
Corresponds to between 3σ and 4σ process level

Capability Targets: Leading-edge lithography aims for Cpk > 1.67 on CD and overlay. Achieving this requires tight process control, predictive maintenance, and continuous improvement. Monthly capability studies verify performance and identify opportunities.

4. Defect Density Reduction

Pareto Analysis (80/20 Rule)

Pareto principle: 80% of defects from 20% of causes. Pareto charts prioritize improvement efforts:

Example Lithography Defect Pareto:
Particles: 450 defects (45% of total)
Incomplete develop: 250 (25%, cumulative 70%)
Resist bridging: 150 (15%, cumulative 85%)
Alignment error: 80 (8%, cumulative 93%)
Other: 70 (7%, cumulative 100%)

→ Focus on top 3 categories for 85% improvement potential

Root Cause Analysis: 5 Whys and Fishbone

5 Whys drills from symptoms to root cause through iterative questioning:

  1. "Why are there particles?" → "Filter efficiency degraded"
  2. "Why degraded?" → "Filter not replaced on schedule"
  3. "Why not replaced?" → "PM schedule not followed"
  4. "Why not followed?" → "PM alert system failed"
  5. "Why failed?" → "System not integrated with CMMS"
  6. Root Cause: Lack of integrated PM tracking

Fishbone (Ishikawa) Diagram categorizes causes into 6M:

8D Problem Solving

Eight Disciplines:
D1: Form cross-functional team
D2: Describe problem (5W2H: Who, What, Where, When, Why, How, How Many)
D3: Implement interim containment
D4: Identify root cause (5 Whys, Fishbone, DOE)
D5: Choose permanent corrective actions
D6: Implement and validate actions
D7: Prevent recurrence (update procedures, training)
D8: Recognize team and close out
Defect Reduction Roadmap: Baseline pareto, analyze top 3 defects, implement corrective actions with DOE validation, update procedures, verify via SPC, repeat for next pareto. Target: 50% defect reduction per year.

5. Failure Mode and Effects Analysis (FMEA)

FMEA Scoring and Risk Priority Number

FMEA proactively identifies potential failures. Each mode scored on Severity (S), Occurrence (O), Detection (D), all on 1-10 scale:

Severity (S): Impact on customer/process
1-3: Minor, 4-6: Moderate, 7-8: High, 9-10: Critical

Occurrence (O): Likelihood of failure
1-2: Remote (< 1/10,000), 9-10: Very High (> 1/10)

Detection (D): Ability to detect before impact
1-2: Almost certain, 9-10: No current controls

Risk Priority Number: RPN = S × O × D (range 1-1000)
Priority: RPN > 100 → action, RPN > 200 → critical

Mitigation Strategies

For high-RPN failure modes, reduce S, O, or D through systematic actions:

FMEA Best Practices: Conduct during process development (Design FMEA) and before production ramp (Process FMEA). Review quarterly or after process changes. Track RPN reduction over time. Cross-functional teams ensure comprehensive identification.

6. Yield Modeling

Poisson Yield Model

Simplest yield model assumes random Poisson-distributed defects:

Y = exp(-D₀ × A)

Y = die yield (fraction good)
D₀ = defect density (defects/cm²)
A = die area (cm²)

Example: D₀ = 0.5/cm², A = 1 cm²
Y = exp(-0.5) = 0.607 = 60.7%

Assumes: Random distribution, all defects fatal, uniform density
Reality: Clustering occurs, model is pessimistic

Negative Binomial (Murphy) Model

Accounts for defect clustering with parameter α:

Y = [1 + (D₀ × A) / α]^(-α)

α → ∞: No clustering (reduces to Poisson)
α = 1: High clustering
α = 2-10: Typical fab processes

Example: D₀ = 0.5, A = 1, α = 2
Y = [1 + 0.5/2]^(-2) = 1.25^(-2) = 0.64 = 64%
Better yield than Poisson due to clustering effect

Critical Area Analysis

Not all defects equally fatal. Critical area A_c(x) is the area where a defect of size x causes failure. Integrate over defect size distribution for accurate yield prediction.

Yield Learning: Fab yield follows Y = Y_max × (1 - exp(-t/τ)), where τ is learning time constant. Mature fabs: 70-90% die yield. New nodes: 10-30% ramping to 60-80% over 12-24 months through systematic defect reduction.

7. Inline Monitoring

Real-Time SPC and Automated Alerts

Modern fabs integrate SPC with MES for real-time control:

Real-Time Architecture:
Metrology tools → Data via SECS/GEM
SPC server → Apply control rules, detect excursions
MES → Log events, trigger alerts, hold lots
Notification → Email/SMS to engineers

Alert thresholds:
Warning (2σ): Email process engineer
Alarm (3σ or WE rule): Email + SMS, lot hold
Critical (multiple violations): Page on-call, escalate

Response time: < 5 minutes from measurement to alert

Fault Detection and Classification (FDC)

FDC monitors equipment sensors for early fault detection:

Inline Monitoring KPIs: Excursion rate (< 1 per 1000 wafers), mean time to detect (< 10 wafers), mean time to resolve (< 4 hours), false alarm rate (< 5%). Continuous improvement minimizes yield impact and downtime.

8. Continuous Improvement

Kaizen Methodology

Kaizen emphasizes small, incremental improvements driven by all employees:

PDCA Cycle (Plan-Do-Check-Act)

  1. Plan: Identify opportunity, analyze current state, develop improvement plan with metrics
  2. Do: Implement on small scale (pilot), collect data, document
  3. Check: Analyze results vs. objectives, statistical comparison, identify gaps
  4. Act: If successful: standardize and deploy. If not: refine and repeat

Six Sigma DMAIC

Define: Charter project, identify CTQ, define scope
Measure: Baseline data, GR&R, calculate current Cpk
Analyze: Root causes (Pareto, Fishbone, hypothesis tests)
Improve: Design solutions (DOE), pilot, verify improvement
Control: Implement SPC, update procedures, train, monitor

Example: Reduce CD variation from Cpk 1.1 to 1.5
Analyze shows 60% variation from dose uniformity, 30% from PEB temp
Improve via DOE optimizes dose map and PEB setpoint → Cpk 1.6
Control via SPC, sustained at 1.6+ for 6 months
CI Culture: Requires leadership support, employee empowerment, training (DOE, SPC, DMAIC), and celebration of results. Typical fab targets: 50-100 improvement projects/year, 20-30% cost reduction over 3-5 years.

9. Sampling Strategies

Sampling Methods

Method Approach Advantages Disadvantages
Random Equal probability for all units Unbiased, valid statistics May miss spatial patterns
Stratified Sample from each stratum Ensures subgroup coverage Requires strata knowledge
Systematic Every kth unit Simple, uniform coverage Can miss periodic variation

Acceptance Sampling (AQL, LTPD)

Single Sampling Plan (n, c):
n = sample size, c = acceptance number
Inspect n units. If defects ≤ c: accept lot. If > c: reject lot.

AQL (Acceptable Quality Level): Producer's risk α = 5%
LTPD (Lot Tolerance Percent Defective): Consumer's risk β = 10%

Example MIL-STD-105E:
Lot 1000 wafers, AQL 1%, Normal Level II
→ Sample n = 80, Acceptance c = 2
≤ 2 defects: accept, ≥ 3 defects: reject

Adaptive Sampling: Adjust frequency based on stability. When Cpk > 1.67 and no violations for 30 days, reduce sampling 50%. When excursion detected, increase to 100% for 20 wafers. Optimizes cost while maintaining risk control.

10. Quality Metrics

First Pass Yield (FPY) and Final Test Yield (FTY)

FPY = Units passing first time / Total units processed

Rolled Throughput Yield (RTY):
RTY = FPY₁ × FPY₂ × ... × FPYₙ

Example (5 steps): 95% × 98% × 99% × 97% × 96% = 85.5%

FTY = Units passing all tests / Units started
Target: FTY > 90% mature, > 60% new nodes

Defect Density Trends

Metric Target (Mature Fab)
Particle Density (> 50nm) < 0.1 per wafer
Pattern Defect Density < 0.5 per cm²
Critical Layer Defect Density < 0.05 per wafer
Defect Reduction Rate 5-10% per month (ramp)

Cpk Trends and Baseline Comparison

Monitor Cpk over time for critical parameters and compare to baseline qualification:

Alert criteria:
- Cpk < 1.33 for 2 consecutive months → Action required
- Cpk decline > 20% month-over-month → Investigation
- Cpk < 1.0 → Immediate escalation, process hold consideration
Quarterly Business Reviews (QBR): Present FPY/FTY trends, Cpk by layer/tool, defect pareto, excursion summary, cost of quality, and improvement initiatives with ROI. Ensures alignment and resources for continuous improvement.