1. The Equipment Commissioning Challenge: Why Smart DOE Is Needed

Before a semiconductor tool ships or enters a customer’s production line, it must undergo a commissioning process — adjusting process parameters (temperature, pressure, flow rate, power, etc.) until the equipment output meets the customer’s process specification requirements.

Traditional commissioning methods suffer from three major problems:

• Excessive test wafer consumption: A full factorial DOE with 5 parameters at 2 levels requires 2^5 = 32 experimental runs. Adding Response Surface Methodology (RSM) center points and axial points, a single round of experiments easily consumes 50-100 test wafers
• Heavy reliance on experience: Senior engineers know “which parameter to adjust first, and by how much,” but this knowledge resides in individuals’ heads and cannot be transferred to new hires or reused across same-type tools
• Extended downtime: Each test wafer requires loading, processing, metrology, and analysis. At most 5-10 experimental runs can be completed per day. Fifty runs translates to 1-2 weeks of equipment downtime

NeuroBox E5200 was designed with a clear objective: find the optimal process recipe that meets specification requirements using the fewest test wafers in the shortest possible time.

2. System Architecture: Five-Stage Closed-Loop Workflow

At its core, the E5200 operates through a five-stage closed-loop workflow, with each stage driven by an independent AI engine:

The system automatically determines “when enough is enough” through its sufficiency assessment — when model accuracy and parameter confidence meet preset criteria, experimentation stops automatically, preventing unnecessary wafer waste.

State Machine Management

Each commissioning process is managed as an independent session with support for interruption and recovery. Session state progresses from creation → DOE generation → data collection → calibration → sufficiency → completion, with fully automated tracking and persistent storage of experimental data and calibration history.

3. Smart DOE Generation: Efficient Parameter Space Coverage

The fundamental problem with traditional DOE is that the number of experimental runs grows exponentially with parameter dimensionality. The E5200 employs a multi-strategy fusion approach to extract maximum information from the fewest possible experiments.

3.1 Space-Filling Design

The primary space-filling strategy is Latin Hypercube Sampling (LHS). LHS uniformly stratifies each parameter dimension, ensuring that experimental points are evenly distributed across the high-dimensional space — far more efficient than both random sampling and full factorial designs.

Consider a 5-parameter CMP commissioning example: a full factorial design requires 32 experimental runs, whereas LHS achieves comparable parameter space coverage with just 10 runs.

3.2 Physics Center Point

In addition to LHS, the system automatically adds the physical center point of the parameter space as a baseline experiment. The center point — where all parameters are set to their midrange values — serves as a “safe” reference experiment and is prioritized for execution, giving engineers a baseline to work from.

3.3 Axial Probes

For each parameter, the system generates 2 axial points (that parameter at its extreme values, all others at center values), enabling rapid identification of each parameter’s main effect. This is essentially a high-efficiency version of One-Factor-At-a-Time (OFAT) experimentation.

3.4 Incremental Design

When additional experiments are needed, the system uses a maximin distance criterion to generate new experimental points — maximizing the minimum distance between the new point and all existing points, ensuring that new experiments explore the most information-scarce regions of the parameter space.

4. Bayesian Calibration Engine: Physics Models + Uncertainty Quantification

The E5200 is not a “black box” model — it is grounded in physics-based semiconductor process models, then applies Bayesian methods to quantify parameter uncertainty.

4.1 Physics Model Calibration

The system includes built-in physics models for nine semiconductor process types (CMP, etch, CVD, PVD, ALD, ion implantation, oxidation, lithography, and diffusion), each with process-specific physical equations.

For CMP, for example, the model is based on the classical Preston equation: removal rate is a function of pressure, rotation speed, flow rate, and temperature. Calibration aims to fit the equation coefficients using experimental data — these coefficients vary by individual tool.

Calibration employs a global optimization algorithm to find the Maximum A Posteriori (MAP) estimate, ensuring the global optimum is found rather than a local minimum.

4.2 Uncertainty Quantification

Simply providing “optimal parameters” is insufficient — the system must also tell engineers “how reliable these parameters are.” The E5200 automatically selects the most appropriate uncertainty quantification method based on sample size:

• Small samples (<10 wafers): Uses Bootstrap resampling. The experimental data is resampled with replacement multiple times, recalibrating each time, and the parameter distribution is computed statistically. This approach is more robust for non-Gaussian posteriors
• Medium samples (≥10 wafers): Uses Laplace approximation. Computes the inverse of the Hessian matrix at the MAP point as the covariance matrix, offering higher computational efficiency

4.3 Parameter Identifiability Assessment

Uncertainty quantification results directly indicate whether parameters are adequately constrained:

4.4 Predictive Uncertainty Propagation

Parameter uncertainty is propagated to prediction outputs through Monte Carlo sampling. The system draws samples from the parameter posterior distribution, evaluates the physics model for each sample, and outputs the prediction mean and 95% confidence interval. Engineers see not just “predicted value = 50 nm,” but rather “predicted value = 50 ± 3 nm (95% confidence).”

5. Active Learning Recommendations: Maximizing the Value of Every Test Wafer

This is the core innovation of the E5200. When the sufficiency assessment determines “more data is needed,” the system does not simply request “run a few more experiments.” Instead, it precisely tells the engineer “what parameter settings the next experiment should use, and why”.

5.1 Information Gain Acquisition Function

The central question in active learning is: which point in the parameter space will most effectively reduce model uncertainty?

The E5200 designs its acquisition function based on the Fisher Information Matrix. For each candidate experimental point, the system calculates: if an experiment is run at this point, how much will the total parameter variance decrease? The point with the highest information gain is selected as the recommendation.

5.2 Efficient Covariance Updates

When recommending multiple experimental points, the system uses an efficient matrix update algorithm to immediately update the parameter covariance matrix after each new point is selected, ensuring subsequent recommendations do not select information-redundant points. This greedy strategy achieves an effective balance between computational efficiency and recommendation quality.

5.3 Recommendation Output

Each recommendation includes:

• Experimental parameters: Specific values for each input parameter
• Expected information gain: A quantified measure of the experiment’s expected value
• Target parameters: Which model parameters this experiment primarily constrains
• Recommendation rationale: A natural-language explanation, e.g., “The pressure coefficient currently has the highest uncertainty; an experiment in the high-pressure region is recommended to improve its constraint”

6. Process Window Analysis and Recipe Optimization

6.1 Virtual DOE and Response Surface

Once the physics model is calibrated, the E5200 can generate a complete response surface through model predictions — without running any additional physical experiments. The system automatically performs grid scans across key parameter pairs, generating 2D response surface plots with uncertainty bands, allowing engineers to visually identify the “sweet spot” in the parameter space.

6.2 Global Sensitivity Analysis

The system automatically analyzes how strongly each input parameter influences the output and ranks them by impact. For example, in CMP commissioning, the system might report: “Pressure has the greatest impact on removal rate (45% contribution), followed by rotation speed (30%), while flow rate has a smaller effect (15%)” — helping engineers focus on the parameters that matter most.

6.3 Inverse Recipe Optimization

Traditional workflow follows “given parameters → observe results.” The E5200 supports inverse optimization — “given target specifications → automatically compute the optimal recipe”:

• Input the customer’s specification limits (LSL/USL)
• A global optimization algorithm searches the parameter space for the recipe that maximizes specification compliance probability
• Outputs the optimal recipe + predicted values + confidence intervals + Cpk

6.4 Process Window Visualization

Starting from the optimal recipe, the system automatically scans the surrounding parameter space to generate process window heat maps:

• Green region: High specification compliance probability (Cpk ≥ 1.33), safe operating zone
• Yellow region: Boundary zone (1.0 ≤ Cpk < 1.33), requires monitoring
• Red region: Does not meet specification (Cpk < 1.0), should be avoided

Engineers learn not only “what the optimal recipe is” but also “how far the recipe can deviate and still remain safe” — critical for managing parameter variations in volume production.

7. Transfer Learning: Faster Commissioning Across Same-Type Tools

This is the key technology enabling the E5200’s “data flywheel.”

7.1 Background

Tools of the same type share the same physics model structure, but their model parameters differ due to individual tool variations (manufacturing tolerances, installation differences, component lot variations, etc.). Traditional approaches commission each tool independently, entirely failing to leverage prior experience.

7.2 Bayesian Transfer Framework

The E5200 employs Bayesian prior transfer:

1. First tool (Machine A): Full commissioning workflow, approximately 15 test wafers. Produces parameter estimates θ_A and covariance matrix Cov_A
2. Second tool (Machine B): Uses θ_A as the prior distribution. Only 2-3 test wafers are needed to combine with the prior and obtain Machine B’s posterior estimate
3. Subsequent tools: The prior strengthens with each additional tool commissioned. By the fifth tool, as few as 3-5 test wafers may suffice

7.3 Transfer Diagnostics

The system automatically detects the significance of parameter shifts. If a tool’s parameters deviate substantially from the prior (potentially indicating a tool anomaly or different configuration), the system issues a warning and recommends independent calibration to prevent erroneous transfer.

8. Data Flywheel: Continuous Self-Improvement

The E5200’s data flywheel effect manifests across multiple levels:

Key Mechanisms:

• Session persistence: All data from each commissioning session (experimental design, metrology results, calibration history, recommendation logs) is automatically saved, with support for interruption recovery and historical review
• Prior accumulation: Calibration results from each tool are automatically incorporated into the transfer learning knowledge base
• Adaptive bounds: The system automatically refines parameter search ranges based on historical calibration results, avoiding wasted experiments in regions already known to be ineffective

9. Deployment and Performance

9.1 Deployment

The E5200 is deployed on an NVIDIA Jetson edge device, sharing the hardware platform with NeuroBox E3200. Core algorithms are implemented using numerical computation libraries and do not require GPU acceleration, running efficiently on edge hardware.

9.2 Typical Commissioning Workflow

9.3 Performance Metrics

9.4 Comparison with Traditional Methods

© 2025-2026 MST Semiconductor (Shanghai) Co., Ltd. · ai-mst.com · This whitepaper is protected by intellectual property rights