Run-to-Run Control: The AI System That Automatically Tunes Every Wafer

What Is Run-to-Run Control and How Does It Differ from Traditional APC?

Semiconductor manufacturing has always relied on Statistical Process Control (SPC) — monitoring process outputs and raising alarms when metrics drift outside control limits. But SPC is fundamentally reactive. By the time an SPC chart signals an out-of-control condition, dozens of wafers may have already been processed with suboptimal recipes.

Run-to-Run (R2R) control is proactive. It adjusts equipment recipe parameters between consecutive wafer runs to compensate for known disturbances — chamber conditioning effects, consumable wear, incoming material variation, and environmental drift. Rather than waiting for a problem to manifest in metrology data, R2R anticipates drift and pre-corrects.

The distinction is significant. Traditional APC (Advanced Process Control) often refers to fault detection and classification (FDC) combined with SPC — essentially sophisticated monitoring. R2R goes further: it closes the control loop by automatically writing recipe adjustments back to the equipment, creating a self-correcting system that maintains process targets without human intervention.

How Do EWMA and Model-Based R2R Algorithms Work?

The most widely deployed R2R algorithm is the Exponentially Weighted Moving Average (EWMA) controller. Its elegance lies in simplicity:

EWMA Control Logic: The controller maintains a running estimate of process bias — the difference between the target value and actual output. After each measured wafer, this estimate updates using a weighted average: new estimate = lambda x (latest error) + (1 – lambda) x (previous estimate). The smoothing factor lambda (typically 0.2-0.4) balances responsiveness against noise sensitivity. The recipe offset for the next run equals the negative of this bias estimate divided by the process gain.

EWMA works well for single-input, single-output processes with linear behavior. But modern semiconductor processes are multi-variable and nonlinear. A CVD process might have 6 controllable parameters (temperature, pressure, gas flows) affecting 3 output metrics (thickness, uniformity, stress) with complex interactions.

Model-Based Predictive R2R: For these multi-variable cases, model-based controllers use process models (linear regression, neural networks, or physics-informed models) to predict how recipe changes will affect all outputs simultaneously. The controller solves an optimization problem: find the recipe adjustment that minimizes predicted deviation from all targets while respecting equipment constraints. This approach handles interactions naturally — it knows that increasing temperature to fix thickness will also affect stress, and compensates accordingly.

MST NeuroBox E3200 implements both EWMA and model-based R2R, automatically selecting the appropriate algorithm based on process complexity and available training data.

What Yield Improvements Can Fabs Expect from R2R?

The yield impact of R2R depends on the process step and baseline control maturity:

Etch Processes: Critical dimension (CD) control via R2R typically reduces CD variation by 40-60%, translating to 2-4% yield improvement at advanced nodes where CD tolerance is measured in sub-nanometer ranges. For a fab running 50,000 wafer starts per month at $5,000 average selling price, a 3% yield improvement generates $7.5 million in additional monthly revenue.

CMP Processes: Removal rate variation is the primary yield limiter in chemical-mechanical planarization. R2R control of polish time and pressure based on incoming film thickness reduces within-lot variation by 50-70%. Fabs report 1-3% yield gains and significant reduction in over-polish and under-polish defects.

CVD/PVD Processes: Film thickness and uniformity control via R2R reduces wafer-to-wafer variation by 30-50%. The compounding effect across multiple deposition steps can yield 2-5% cumulative improvement.

Overall Impact: A comprehensive R2R deployment across 15-20 critical process steps typically delivers 5-10% aggregate yield improvement. At a mid-size 300mm fab, this represents $20-50 million in annual value — making R2R one of the highest-ROI investments available to fab operations teams.

Why Does Closed-Loop Automation Matter for Process Engineers?

Process engineers at a typical fab spend 30-40% of their time on manual recipe tuning — reviewing metrology data, calculating offsets, entering recipe changes, and verifying results. This is skilled, repetitive work that scales linearly with the number of tools and products.

R2R automation eliminates this manual loop. The system continuously monitors, calculates, and adjusts — 24 hours a day, 7 days a week, with consistent mathematical precision. Engineers shift from tuning recipes to managing control strategies, analyzing trends, and improving process capability.

The consistency benefit is equally important. Manual tuning varies with engineer experience, shift patterns, and workload. One engineer might adjust aggressively; another conservatively. R2R applies the same optimal control law every time, eliminating human variability that can account for 10-20% of total process variation.

For equipment makers, embedding R2R capability into their tools creates a powerful differentiation story. Instead of selling a machine that requires expert tuning, they deliver a self-optimizing system that maintains spec from day one — reducing the customer burden of supporting the equipment and increasing reorder likelihood.

What Data Infrastructure Does R2R Require?

R2R is only as good as its data pipeline. The system needs:

Real-Time Equipment Data: Recipe parameters and sensor traces via SECS/GEM or EDA (Equipment Data Acquisition) interfaces, with latency under 500 milliseconds. The controller must know exactly what recipe ran and what the equipment did during that run.

Metrology Feedback: Actual measurement values linked to specific wafer runs, typically arriving 1-4 hours after processing (the metrology queue delay). The R2R system must correctly associate measurements with the runs that produced them — a data alignment challenge that trips up many implementations.

Context Metadata: Lot ID, wafer slot, product, layer, recipe version, chamber ID, PM cycle count. Without this context, the controller cannot distinguish between a process drift (requiring correction) and a product change (requiring a different target).

Recipe Write-Back: The ability to programmatically modify equipment recipe parameters — the most sensitive integration point. This requires tight coordination with the Manufacturing Execution System (MES) and equipment communication protocols. Security controls must ensure the R2R system can only modify designated parameters within pre-approved ranges.

MST NeuroBox provides a unified data infrastructure layer that handles collection, alignment, and write-back across equipment from multiple vendors — solving the integration challenge that is often the biggest barrier to R2R deployment.

How Should Fabs Deploy R2R for Maximum Impact?

Successful R2R deployment follows a priority-driven approach:

Step 1 — Target Selection: Rank process steps by yield sensitivity multiplied by current process capability index (Cpk). Steps with high yield sensitivity and low Cpk are the best R2R candidates. Typically, this means critical etch and CMP steps at advanced nodes.

Step 2 — Baseline and Model: Collect 4-8 weeks of production data with deliberate recipe variations (small DOE overlays) to characterize process gain and dynamics. Build the process model and validate prediction accuracy. This phase costs 200-500 wafers but provides the foundation for the entire control system.

Step 3 — Shadow Mode: Run R2R in advisory mode for 2-4 weeks. The system calculates recipe adjustments but does not execute them. Engineers review recommended changes against their manual adjustments. This builds trust and identifies edge cases.

Step 4 — Closed-Loop Activation: Enable automatic recipe write-back with conservative guardrails — small maximum adjustment per run, limited total cumulative offset, automatic lockout if predictions diverge from measurements. Gradually widen guardrails as confidence grows.

Step 5 — Expansion: Roll out to additional process steps, products, and tool groups. Implement cross-tool control where one tool R2R adjustments account for upstream process variation.

The timeline from project start to first closed-loop operation is typically 3-4 months, with full fab-wide deployment taking 12-18 months. The investment pays back within the first 6 months of closed-loop operation for most fabs.