1. Why Edge AI: The Production Data Dilemma

A semiconductor production line generates gigabytes of sensor data every second, yet the utilization rate of this data remains extremely low. In traditional architectures, data must travel from equipment to EAP to database to analytics platform — layer by layer — introducing latencies ranging from minutes to hours. By the time engineers see the analysis results, dozens of wafers may have already been processed with suboptimal parameters.

NeuroBox E3200 is built on the principle of “bringing AI to where the data is generated”:

• Zero-latency data acquisition: Collects data directly from the equipment port via SECS/GEM protocol, bypassing the EAP middleware layer
• Real-time inference: AI models run on Jetson hardware at the equipment level, completing predictions in 35-50 ms
• Data stays on-premises: All computation is performed locally, meeting the semiconductor industry’s stringent data security requirements
• Closed-loop control: VM predictions directly drive R2R parameter compensation with no manual intervention required

2. System Architecture: Three Software Layers + Edge Hardware

The E3200 software architecture is organized into three independent packages, each serving a distinct function:

Hardware Platform

• Processor: NVIDIA Jetson Orin NX (8-core ARM + GPU)
• Inference acceleration: Deeply optimized edge inference engine, approximately 50x faster than generic frameworks
• Operating system: Linux / JetPack
• Interfaces: Gigabit Ethernet (HSMS communication), USB (debug), HDMI (local display)

3. SECS/GEM Communication Stack: Direct Equipment Data Acquisition

The E3200 includes a complete SECS/GEM protocol stack, compliant with SEMI E5 (SECS-II message format) and SEMI E37 (HSMS transport protocol) standards.

3.1 HSMS Protocol Implementation

• Connection modes: Supports both Active (initiates connection to equipment) and Passive (waits for equipment to connect)
• Message types: Full message family support including DATA_MESSAGE, SELECT, DESELECT, LINKTEST, SEPARATE, and more
• Heartbeat mechanism: Default 60-second Linktest interval with full T3-T8 timeout parameter configuration
• Asynchronous I/O: Non-blocking communication architecture supporting concurrent transactions on a single connection
• Maximum message size: Supports bulk Trace Data transmission

3.2 SECS-II Encoder/Decoder

Full support for all data types defined in the SECS-II specification (List, Binary, Boolean, ASCII, integer, floating-point, etc.), with automatic type inference and bidirectional type conversion.

3.3 Equipment Configuration Driver

The E3200 adapts to different equipment through configuration files — one codebase supports all equipment types, eliminating the need for custom code per tool. Configuration covers equipment identification, connection parameters, data variable mapping, remote command definitions, and event-report link configuration.

The system has been validated for compatibility with equipment from Applied Materials, Lam Research, Tokyo Electron, Axcelis, and other major semiconductor equipment vendors across multiple tool types.

4. Four-Layer VM Prediction Architecture: From Physics to Adaptation

This is the core technical innovation of NeuroBox E3200. Traditional VM approaches typically rely on a single model (e.g., PLS or a machine learning model), but the complexity of semiconductor processes demands a more refined modeling strategy. The E3200 employs a four-layer cascaded prediction architecture:

4.1 L2: Physics-Based Operator

First-principles modeling based on semiconductor process physics, with dedicated physical formulas built in for each process type:

For nine semiconductor processes (ion implantation, etch, CMP, CVD, PVD, ALD, oxidation, lithography, and diffusion), the E3200 includes first-principles physics-based prediction models. For example, CMP is modeled using the Preston equation, and oxidation uses the Deal-Grove model — these classical physics models have been engineered for efficient execution on edge devices.

The key advantage of physics-based models is that they do not require large datasets — they are grounded in process fundamentals rather than statistical fitting, enabling reasonable predictions from first deployment.

4.2 L3: Residual Correction Network

Physics models cannot perfectly describe real-world processes (equipment-specific variations, unmodeled interaction effects, etc.). The L3 network learns the residual between the physics model prediction and actual metrology values:

y_residual = y_actual – y_physics
y_prediction = L2_physics(x) + L3_residual(x)

• Network architecture: Lightweight multi-layer neural network, optimized for edge deployment with ultra-low inference latency
• Machine Embedding: One embedding vector per tool, capturing equipment-specific variations
• Training objective: Learns residuals rather than absolute values, reducing data requirements
• Accuracy: With L3 enabled, VM accuracy improves significantly (MAPE reduction exceeding 50%)

4.3 L4: RLS Online Learning (Real-Time Drift Tracking)

Equipment state drifts over time (consumable wear, component aging), causing L3 model predictions to gradually deviate. The L4 layer uses Recursive Least Squares (RLS) for real-time drift tracking:

• Adaptive forgetting: Exponentially decays old data weights, automatically adapting to equipment state changes
• Online update: Millisecond-level model weight updates upon receipt of each wafer’s metrology feedback
• Numerical stability: Industrial-grade numerical stability safeguards for reliable long-term operation
• State management: Supports checkpoint save/load; state can be reset after PM events

4.4 UQ: Uncertainty Quantification

Providing a prediction alone is not sufficient — the system must also convey “how reliable this prediction is”:

• Model ensemble: Multiple independently trained models each generate predictions; the mean and standard deviation are computed
• Confidence interval: Outputs y_mean ± 2σ for a 95% confidence interval
• Low-confidence alerts: When σ exceeds a threshold, an alert is automatically triggered with a recommendation for physical metrology
• Optimizer integration: σ directly influences the R2R optimizer’s trust region radius — higher uncertainty automatically results in more conservative optimization steps

5. Constrained Optimizer and R2R Closed-Loop Control

5.1 Constrained Optimization Problem

The core of R2R control is computing the parameter compensation for the next run based on VM predictions. The E3200 formulates this as a constrained optimization problem:

5.2 Solver Implementation

• Convex optimization solver: Efficient quadratic programming solver with millisecond-level solve times
• Gradient computation: Computes true gradients from the actual VM engine
• Adaptive trust region: Higher UQ uncertainty automatically yields more conservative step sizes
• Safety checks: Parameter whitelist verification, interlock validation, and step-size upper-bound safeguards

5.3 VM-R2R Closed Loop

The complete closed-loop control flow:

1. Equipment reports the current run’s process parameters via SECS/GEM
2. The VM engine outputs predicted metrology values and confidence levels within 35 ms
3. The optimizer computes parameter compensation based on predicted and target values
4. Compensation values are sent to the equipment via the SECS/GEM S2F41 command
5. The next wafer is processed with the compensated parameters
6. Actual metrology values are fed back to the L4 layer for model updates

The entire closed loop completes within 50 ms with no engineer intervention required.

6. FDC: Fault Detection and Classification

Unlike traditional threshold-based alarm systems, the E3200’s FDC module employs a dual-engine architecture:

6.1 Supervised Classification Engine

• Training data: Historical wafer process parameters + yield labels
• Feature engineering: Automatically extracts multi-dimensional statistical features from Trace Data
• Output: Per-wafer fault probability + feature importance ranking (Top-N critical parameters)
• Advantage: Highly interpretable — engineers can directly see which parameters are anomalous

6.2 Unsupervised Anomaly Detection

• Training data: Trained exclusively on normal wafer data (no fault labels required)
• Detection principle: Deep learning-based anomaly pattern recognition
• Advantage: Capable of detecting unknown fault modes without relying on historical fault samples

6.3 Particle Tracking (Particle Analysis)

A dedicated analysis module for particle defects:

• Event chain analysis: Traces the timeline of particle generation, correlating equipment operation events
• Multi-layer analysis architecture: Physical layer → Semiconductor layer → Statistical layer → ML layer, progressively deepening the analysis
• Time-series anomaly detection: Time-series model-based detection of abrupt changes in particle counts

7. Equipment Intelligence Platform (EIP)

The EIP (Equipment Intelligence Platform) is the upper-level application layer of the E3200, integrating outputs from VM, R2R, and FDC to provide unified equipment intelligence management:

7.1 Equipment Health Scoring

Comprehensively evaluates equipment state across multiple dimensions, outputting a health index from 0 to 100. Dimensions include sensor drift, actuator response, seal integrity, temperature control precision, and more. Health trends are visualized with early degradation warnings.

7.2 PM Prediction

Based on equipment operational data and health trends, predicts Remaining Useful Life (RUL) and intelligently recommends optimal PM time windows.

7.3 AI Diagnostic Agent

When FDC raises an alert or health scores decline, the AI Diagnostic Agent automatically analyzes the situation:

• Cross-references the historical fault knowledge base to recommend the most probable root causes
• Cross-validates anomaly timelines across multiple parameters
• Outputs structured diagnostic reports with recommended corrective actions

7.4 Alarm Bridge

Correlates native equipment alarms with E3200 AI analysis results to provide enriched alarm context: not just “temperature exceeded limit,” but rather “temperature exceeded limit + 7-day drift trend + correlated RF power variation + recommended inspection of Heater Zone 3 thermocouple.”

7.5 Web Dashboard

Access the E3200 local management interface directly through a browser:

• Real-time monitoring dashboard (equipment status, VM predictions, FDC analysis)
• R2R control panel (parameter compensation history, optimization trajectory)
• Historical data charts and visualizations
• Real-time log streaming

8. Data Flywheel: Continuous Self-Improvement

The E3200’s four-layer architecture inherently enables continuous evolution:

Key Mechanisms:

• L3 periodic retraining: Automatically triggered with every 1,000 new wafer data points or after PM events
• L4 real-time updates: Millisecond-level weight updates upon each wafer’s metrology feedback
• FDC model evolution: FDC models automatically update after engineers confirm each root cause
• Cross-tool transfer: L3 models trained on one tool of the same type can be transferred to another, with Machine Embedding automatically adapting to individual tool variations

9. Deployment and Performance

9.1 Deployment Procedure

• Connect the E3200 hardware unit to the equipment’s HSMS port via Ethernet
• Upload the equipment configuration file (IP, port, variable mapping)
• The system automatically completes HSMS connection → S1F13 communication establishment → event configuration → data collection
• L2 physics model is immediately available; L3/L4 come online automatically as data accumulates

9.2 Performance Metrics

9.3 SEMI Standards Compliance

• SEMI E5: SECS-II message format
• SEMI E37: HSMS high-speed transport
• SEMI E30: GEM (Generic Equipment Model)
• SEMI E40: Process management

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