Wafer Uniformity Control: The Dual Challenge of W2W and WIW

Author: MST Semiconductor | Category: Semiconductor Process Control | Keywords: Wafer Uniformity, W2W, WIW, Semiconductor Process Control

1. W2W: Inter-Wafer Consistency — Making Every Wafer the Same

W2W uniformity focuses on the consistency of process results across different wafers. Ideally, when processing 100 consecutive wafers, the average film thickness (or CD, etch depth, etc.) on every wafer should be identical.

1.1 Sources of W2W Variation

The primary factors causing inter-wafer variation include:

• Process drift: As the number of processed wafers increases, chamber wall deposition accumulates and consumables wear, causing process conditions to slowly change. This is the most common source of W2W variation.
• Lot effects: Wafers from different lots may come from different upstream processes, resulting in variation in incoming wafer conditions.
• Equipment state fluctuations: Random disturbances such as minor temperature control lag and short-term gas flow fluctuations.
• PM impact: Equipment characteristics undergo step changes after preventive maintenance, requiring re-stabilization.

1.2 Traditional W2W Control Methods

Traditionally, W2W control relies on periodic measurement and manual compensation. Engineers sample one wafer at regular intervals, and upon detecting drift, manually adjust recipe parameters. A more advanced approach introduces Run-to-Run (R2R) controllers that automatically adjust parameters based on algorithms such as EWMA (Exponentially Weighted Moving Average).

However, traditional R2R has its limitations: the linear model assumption becomes inaccurate for large drifts; single-variable control cannot handle multi-parameter coupling; and controller parameters (such as gain and filter coefficients) require manual tuning.

2. WIW: Intra-Wafer Uniformity — Making Every Point on Each Wafer the Same

WIW uniformity focuses on consistency across different positions within a single wafer. It is typically quantified using %Range or %1-sigma, with a typical requirement of film thickness uniformity < 2% (1-sigma).

2.1 WIW Non-Uniformity Patterns

WIW non-uniformity typically manifests as specific spatial patterns. Common patterns include:

• Center-to-Edge variation: The most common pattern, presenting as thicker film at the wafer center and thinner at the edge (or vice versa). This is usually related to gas flow field distribution, temperature gradients, and plasma density distribution.
• Left-Right asymmetry: Systematic differences between one side of the wafer and the other, typically indicating equipment hardware symmetry issues (such as showerhead tilt or exhaust port position offset).
• Ring patterns: A ridge or trough appearing at a certain radius, possibly related to focus ring edge effects or temperature zone boundaries.
• Random noise: Fluctuations with no apparent spatial pattern, usually related to particle contamination or measurement noise.

2.2 Traditional WIW Control Methods

WIW optimization is primarily achieved through equipment parameter adjustment:

• Adjusting showerhead height (gap) to modify gas distribution
• Adjusting multi-zone temperature control (e.g., center/middle/edge three-zone heaters) to modify the temperature field
• Adjusting pressure and gas flow rates to influence reactant distribution
• On some tools, adjusting plasma antenna power distribution

The challenge with traditional methods is that each parameter’s effect on uniformity is highly complex, and changes often produce cascading effects — improving center-to-edge variation may introduce left-right asymmetry.

3. The Relationship and Conflicts Between W2W and WIW

W2W and WIW may appear to be two independent dimensions, but in practical process control, they have subtle interconnections and even conflicts:

3.1 Coupled Control Dimensions

Many equipment parameters affect both W2W and WIW simultaneously. For example, increasing gas flow in a CVD process can compensate for W2W drift (maintaining average film thickness), but may change the gas distribution pattern inside the chamber, degrading WIW.

3.2 Measurement Frequency Trade-Offs

W2W control requires per-wafer or per-lot measurement feedback, but typically only a few points (5 or 9) are measured to obtain the mean. WIW analysis requires dense multi-point measurements (49 points or more), but this cannot be done for every wafer — measurement itself is a throughput bottleneck.

3.3 Different Time Scales

W2W variation is short-term “wafer-to-wafer” change requiring per-wafer or per-lot control. WIW spatial patterns typically change more slowly (related to equipment hardware condition), constituting a medium-to-long-term problem. The two require different control cadences and strategies.

3.4 Conflicting Optimization Objectives

Sometimes, the parameter settings that maximize W2W consistency are not optimal for WIW. For example, a particular temperature setting may best compensate for lot-to-lot drift, but that temperature produces a larger radial temperature gradient, yielding suboptimal intra-wafer uniformity. Engineers must find the right balance between the two.

4. How AI Optimizes Both Dimensions Simultaneously

Traditional methods treat W2W and WIW as two independent problems and address them separately — which is precisely the root cause of conflicts. AI’s core advantage lies in unified modeling and joint optimization.

4.1 Spatial-Temporal Joint Modeling

The AI system simultaneously models both the temporal evolution of process results (W2W dimension) and their spatial distribution (WIW dimension). Model inputs include recipe parameters, equipment sensor data, and upstream process information. The output is not just a prediction of the wafer mean, but a prediction of the complete wafer spatial distribution.

When the model recommends adjusting a parameter to compensate for W2W drift, it simultaneously evaluates that adjustment’s impact on WIW. If the impact is unacceptable, the model automatically searches for alternatives.

4.2 Multi-Objective Pareto Optimization

The AI optimizer searches for Pareto-optimal solution sets in the parameter space — i.e., solutions that improve WIW without degrading W2W, or vice versa. Engineers can select the appropriate operating point on the Pareto front based on current production priorities.

For example, on yield-sensitive critical layers, WIW uniformity may take priority; on less yield-sensitive layers, WIW requirements can be relaxed somewhat in exchange for better W2W consistency.

4.3 Hierarchical Control Architecture

The recommended AI control architecture employs a hierarchical strategy:

• Fast layer (per-wafer/per-lot): The AI-R2R controller handles W2W compensation, primarily adjusting parameters that do not affect WIW (such as process time, power, and other “uniform adjustment” parameters).
• Medium layer (daily/per-shift): WIW trends are assessed based on multi-point measurement data, with adjustments made to spatial distribution parameters (such as temperature zoning, gap, etc.) as needed.
• Slow layer (weekly/post-PM): A comprehensive equipment state evaluation is performed, the model is updated, and baseline parameters are re-optimized.

The three layers work in concert, ensuring both real-time W2W correction capability and long-term WIW stability.

4.4 Virtual Metrology Support

Dense WIW measurements cannot be performed on every wafer, but AI-based Virtual Metrology (VM) can predict each wafer’s spatial distribution based on equipment sensor data. This way, even lots with only 5-point measurements can obtain complete spatial information through virtual metrology, providing high-frequency feedback for WIW control.

5. Summary

Wafer uniformity control is an enduring challenge in semiconductor manufacturing. W2W and WIW each present their own difficulties, and more critically, the coupling and conflicts between them require systematic treatment. AI’s advantage lies in its ability to build unified multi-dimensional models, find optimal operating points in complex parameter spaces that simultaneously satisfy multiple uniformity metrics, and achieve real-time, adaptive optimization through hierarchical control architectures.