产品概述 Product Overview
NeuroBox™ E5200 是迈烁集芯自主研发的半导体设备 AI 边缘智能平台。它是一个工业级无风扇边缘计算盒子,通过 SECS/GEM 协议直连半导体生产设备,实现虚拟量测(VM)、闭环自动调机(R2R)、设备智能诊断(EIP)三大核心能力。
核心定位
一个NeuroBox,三大能力:
- VM 虚拟量测 — 用 AI 预测晶圆质量指标,减少 10-50% 破坏性检测
- R2R 闭环调机 — 输入客户 spec,直接输出最优 recipe,实现闭环自动优化
- EIP 设备诊断 — 多模态故障诊断 + 健康评分 + 预测性维护
NeuroBox™ E5200 is an AI edge intelligence platform for semiconductor equipment, independently developed by MST. It is an industrial-grade fanless edge computing device that connects directly to semiconductor production equipment via SECS/GEM protocols, delivering three core capabilities: Virtual Metrology (VM), Run-to-Run Auto-Tuning (R2R), and Equipment Intelligence Platform (EIP).
Core Positioning
One NeuroBox, three capabilities:
- VM Virtual Metrology — AI-powered quality prediction, reducing destructive inspections by 10-50%
- R2R Closed-Loop Tuning — Input customer specs, output optimal recipe with closed-loop optimization
- EIP Equipment Diagnostics — Multi-modal fault diagnosis + health scoring + predictive maintenance
系统界面与演示 System Interface & Demo
产品演示视频
Product Demo Video
NeuroBox™ E5200 产品演示 NeuroBox™ E5200 Product Demo
行业痛点:传统工艺开发为什么这么慢? Industry Challenge: Why Is Traditional Process Development So Slow?
半导体设备商在交付设备时,需要为客户完成工艺配方(recipe)的开发与验证。传统方法依赖工程师经验,逐一调整参数、反复试片测量,效率低下且高度依赖个人经验。
典型场景
客户要求:"CMP 工艺去除速率 200±20 nm/min,均匀性 WIWNU < 5%"
传统做法:工程师凭经验选参数 → 跑 1 片 → 测量 → 不达标 → 改参数 → 再跑 → 再测…… 循环 30-50 次。
| 传统方法 | NeuroBox™ AI 方案 | |
|---|---|---|
| 实验晶圆数量 | 30-50 片 | 10-15 片 |
| 开发周期 | 2-4 周 | 3-5 天 |
| 经验依赖 | 强,换人就丢 | 弱,知识沉淀在模型 |
| 可解释性 | 差,不知道为什么能用 | 强,每个参数有物理含义 |
| 多机台部署 | 每台从零开始 | 迁移学习,2-3 片即可 |
| 过程能力保障 | 无法量化 | 输出 Cpk 与满足 spec 概率 |
When semiconductor equipment vendors deliver equipment, they need to develop and verify process recipes for customers. Traditional methods rely on engineer experience, manually adjusting parameters one at a time, running wafers, and measuring — a slow process highly dependent on individual expertise.
Typical Scenario
Customer requirement: "CMP removal rate 200±20 nm/min, uniformity WIWNU < 5%"
Traditional approach: Engineer picks parameters → runs 1 wafer → measures → misses target → adjusts → runs again… repeat 30-50 times.
| Traditional Method | NeuroBox™ AI Solution | |
|---|---|---|
| Experiment Wafers | 30-50 wafers | 10-15 wafers |
| Development Cycle | 2-4 weeks | 3-5 days |
| Experience Dependency | High, lost with turnover | Low, knowledge in models |
| Explainability | Poor, "it works but we don't know why" | Strong, physics-based parameters |
| Multi-Tool Deployment | Start from scratch each time | Transfer learning, 2-3 wafers |
| Process Capability | Cannot quantify | Outputs Cpk & spec conformance probability |
核心技术一:VM 虚拟量测 4 层架构 Core Technology 1: 4-Layer Virtual Metrology Architecture
NeuroBox 采用业界首创的 4 层 VM 架构,将物理先验知识与数据驱动 AI 深度融合,在小样本场景下即可实现高精度质量预测。
基于 Preston 方程、SRIM 表、Arrhenius 动力学等经典物理模型,提供初始预测基线。覆盖离子注入、刻蚀、PVD、CVD、CMP、ALD、热氧化、光刻等 10+ 种工艺类型。
使用 PyTorch 训练的残差神经网络,学习物理模型的系统性偏差。模型仅 82KB,在设备端高效运行。
递推最小二乘法(RLS),实时补偿设备漂移。遗忘因子 λ=0.98,持续适应无需重新训练。
5 模型集成 + MC Dropout 不确定性估计。当预测置信度低时,自动触发实际量测,避免漏检。
关键性能指标
| 预测精度 | MAPE < 6%(离子注入验证) |
| 推理延迟 | < 10 ms / 片 |
| 模型总大小 | < 200 MB |
| 支持工艺数 | 10+ 种(注入、刻蚀、PVD、CVD、CMP 等) |
NeuroBox features an industry-first 4-layer VM architecture that deeply fuses physics-based domain knowledge with data-driven AI, achieving high-accuracy quality prediction even with small sample sizes.
Based on classical physics models (Preston equation, SRIM tables, Arrhenius kinetics), providing initial prediction baselines. Covers 10+ process types including implant, etch, PVD, CVD, CMP, ALD, thermal oxidation, and lithography.
PyTorch-trained residual neural network that learns systematic biases in the physics model. Only 82KB, runs efficiently on-device.
Recursive Least Squares (RLS) for real-time equipment drift compensation. Forgetting factor λ=0.98, continuously adapts without retraining.
5-model ensemble with MC Dropout uncertainty estimation. Automatically triggers physical measurement when prediction confidence is low.
Key Performance Metrics
| Prediction Accuracy | MAPE < 6% (validated on implant) |
| Inference Latency | < 10 ms per wafer |
| Total Model Size | < 200 MB |
| Supported Processes | 10+ (Implant, Etch, PVD, CVD, CMP, etc.) |
核心技术二:快速工艺开发 — 从 spec 到 recipe 的闭环 Core Technology 2: Rapid Process Development — From Spec to Recipe
NeuroBox 的 R2R 系统实现了从 "客户 spec 输入 → 最优 recipe 输出" 的完整闭环,这是传统 APC 方案无法做到的。
生成实验方案
系统自动生成优化的采样方案:"请按以下参数跑 10 片晶圆"——用最少的晶圆数量,全面覆盖参数空间(压力 psi / 转速 rpm / 流量 / 温度等),而非传统一次一因素逐一试错。
跑片 & 建模
录入实验数据后,系统自动建立工艺物理模型,并实时报告每个参数的建模状态——哪些已经准了,哪些还需要更多数据。
智能推荐
系统智能推荐下一片最有价值的实验参数组合,帮你最快找到最优方案,而非盲目试错。
达标完成
输入客户工艺要求(如 "去除速率 200±20 nm/min,均匀性 < 5%"),系统直接输出最优 recipe,并给出满足要求的概率和过程能力 Cpk 值。
过程窗口验证
自动运行虚拟实验,生成 3D 响应曲面和过程窗口热力图,直观展示哪些参数组合最稳定、最可靠。
Smart DOE 技术原理 — 为什么只需 10~15 片晶圆?
传统工艺开发依赖工程师经验逐一调参,50~100 片起步。NeuroBox 内置的 Smart DOE + 贝叶斯标定 + 主动学习 三级闭环,从算法层面根本性地减少了所需实验数量。
第一级:Latin Hypercube Sampling (LHS) — 智能实验设计
假设工艺有 5 个可调参数(压力、转速、时间、流量、温度),传统全因子实验需要 35 = 243 组。LHS 将每个参数的取值范围切成 N 等份,保证每个维度上每个区间恰好被采样一次——类似"数独"原理,仅需 10~15 组实验即可均匀覆盖整个 5 维参数空间。
此外,系统自动在实验方案中嵌入中心点(最安全的起始实验)和可选轴向点(识别单参数主效应),确保实验方案既高效又可靠。
第二级:贝叶斯标定 — 从数据中学习物理规律
实验数据回传后,系统基于物理方程(如 CMP 的 Preston 方程 RR = Kp · Pα · Vβ)进行贝叶斯参数拟合:
- 小样本(< 10 片):使用 Bootstrap 重采样,对非高斯后验更稳健
- 样本充足(≥ 10 片):使用 Laplace 近似(Hessian 矩阵求逆),速度更快
每次标定后,系统自动输出每个模型参数的置信区间和可辨识性评级(充分/中等/不足),工程师一目了然。
第三级:Fisher Information 主动学习 — "下一片跑什么"
这是 NeuroBox 区别于所有传统 DOE 工具的核心差异化能力。
系统计算每个候选实验点对模型参数的 Fisher Information(信息量),数学上精确衡量"跑这组参数能减少多少不确定性"。然后通过 Woodbury(Sherman-Morrison)公式快速更新协方差矩阵,贪心选择信息增益最大的实验点。
直观理解:系统不是问"哪里预测不准"(那是被动学习),而是问"跑哪组参数能最快让整个模型变准"——这才是工艺开发(建模)的正确目标。
完整闭环:迭代至收敛
LHS 生成初始方案(10 组)
均匀覆盖参数空间,客户去机台跑片
贝叶斯标定
拟合物理模型,量化每个参数的不确定性
充分性评估
R² ≥ 0.90 且所有参数 CV < 25%?达标则完成;否则继续
主动学习推荐
Fisher Information 精确计算最有价值的下一组实验,回到 B
Smart DOE vs 传统方法对比
| 传统 OFAT | 全因子 DOE | NeuroBox Smart DOE | |
|---|---|---|---|
| 实验设计 | 逐一调参 | 固定网格 | LHS 均匀覆盖 |
| 所需晶圆(5 参数) | 50~100 片 | 243 片(3 水平) | 10~15 片 |
| 建模方法 | 无模型 | 线性回归 | 物理方程 + 贝叶斯标定 |
| 不确定性量化 | 无 | 无 | 每参数置信区间 + 可辨识性评级 |
| 下一步推荐 | 靠经验 | 无 | Fisher Information 主动学习 |
| 迭代闭环 | 手动 | 一次性 | 自动迭代至数据充分 |
| 计算硬件 | N/A | PC 软件 | 嵌入设备(Jetson NX 边缘) |
三大杀手级功能
反向 Recipe 推荐
输入客户 spec → 直接输出最优 recipe + 满足概率 97.3% + Cpk 1.42。传统 APC 方案无法实现的能力。
机台间迁移学习
机台 A 标定 15 片 → 机台 B 仅需 2-3 片。10 台同型设备传统需 150 片,我们仅需 42 片,节省 72% 晶圆。
嵌入式实时优化
不同于传统的离线仿真工具或独立 APC 软件,NeuroBox 直接嵌入设备控制系统,通过 SECS/GEM 实时闭环。
R2R 技术参数
| 优化算法 | CVXPY + OSQP 约束凸优化 |
| 求解延迟 | < 50 ms / 次 |
| 安全机制 | Shadow 模式、自动回滚、联锁验证 |
| 优化模式 | 逐片 / 逐批 / 手动+自动混合 |
NeuroBox's R2R system achieves a complete closed loop from "customer spec input → optimal recipe output", something traditional APC solutions cannot do.
Generate Experiment Plan
System automatically generates an optimized sampling plan: "Run 10 wafers with these parameters" — covering the full parameter space (pressure psi / velocity rpm / flow / temp) with minimal wafers.
Run Wafers & Build Model
After entering experimental data, the system automatically builds the process physics model and reports real-time status — which parameters are calibrated and which need more data.
Smart Recommendation
System intelligently recommends the most valuable next experiment parameters, helping you find the optimal solution fastest instead of blind trial-and-error.
Meet Spec
Input customer process requirements (e.g., "removal rate 200±20 nm/min, uniformity < 5%") and the system outputs the optimal recipe with spec conformance probability and Cpk values.
Process Window Validation
Automatically runs virtual experiments generating 3D response surfaces and process window heatmaps, showing which parameter combinations are most stable and reliable.
Smart DOE Technical Principles — Why Only 10-15 Wafers?
Traditional process development relies on engineers manually tuning parameters one at a time, typically requiring 50-100 wafers. NeuroBox's built-in Smart DOE + Bayesian Calibration + Active Learning three-tier closed loop fundamentally reduces the number of experiments needed at the algorithm level.
Tier 1: Latin Hypercube Sampling (LHS) — Smart Experiment Design
Consider a process with 5 tunable parameters (pressure, velocity, time, flow, temperature). A traditional full factorial design requires 35 = 243 runs. LHS divides each parameter's range into N equal intervals, ensuring each interval is sampled exactly once per dimension — similar to a "Sudoku" principle. Only 10-15 experiments are needed to uniformly cover the entire 5-dimensional parameter space.
Additionally, the system automatically embeds center points (safest starting experiment) and optional axial points (for identifying single-parameter main effects), ensuring experiment plans are both efficient and reliable.
Tier 2: Bayesian Calibration — Learning Physics from Data
After experimental data is collected, the system fits physics equations (e.g., CMP's Preston equation RR = Kp · Pα · Vβ) using Bayesian parameter estimation:
- Small samples (< 10 wafers): Bootstrap resampling, more robust for non-Gaussian posteriors
- Sufficient samples (≥ 10 wafers): Laplace approximation (Hessian matrix inversion), faster computation
After each calibration, the system automatically outputs confidence intervals and identifiability ratings (well-constrained / moderate / poorly-constrained) for every model parameter.
Tier 3: Fisher Information Active Learning — "What to Run Next"
This is NeuroBox's core differentiator from all traditional DOE tools.
The system computes the Fisher Information of each candidate experiment point with respect to model parameters, mathematically quantifying "how much uncertainty would this experiment reduce." It then uses the Woodbury (Sherman-Morrison) formula for fast covariance matrix updates, greedily selecting the point with maximum information gain.
Intuitive understanding: the system doesn't ask "where is the prediction inaccurate" (passive learning), but rather "which experiment will make the entire model converge fastest" — the correct objective for process development (model identification).
Complete Closed Loop: Iterate to Convergence
LHS Generates Initial Plan (10 runs)
Uniformly covers parameter space; customer runs wafers on the tool
Bayesian Calibration
Fit physics model; quantify uncertainty for each parameter
Sufficiency Assessment
R² ≥ 0.90 and all parameter CV < 25%? If yes, done; otherwise continue
Active Learning Recommendation
Fisher Information precisely identifies the most valuable next experiment; loop back to B
Smart DOE vs Traditional Methods
| Traditional OFAT | Full Factorial DOE | NeuroBox Smart DOE | |
|---|---|---|---|
| Experiment Design | One-at-a-time | Fixed grid | LHS uniform coverage |
| Wafers (5 params) | 50-100 | 243 (3 levels) | 10-15 |
| Modeling | No model | Linear regression | Physics eq. + Bayesian calibration |
| Uncertainty Quantification | None | None | Per-parameter CI + identifiability |
| Next Step Guidance | Engineer intuition | None | Fisher Information active learning |
| Iterative Closed Loop | Manual | One-shot | Auto-iterate until sufficient |
| Compute Hardware | N/A | PC software | Embedded (Jetson NX edge) |
Three Killer Features
Inverse Recipe Optimization
Input customer spec → output optimal recipe + 97.3% conformance probability + Cpk 1.42. A capability traditional APC solutions cannot match.
Tool-to-Tool Transfer Learning
Tool A calibrated with 15 wafers → Tool B needs only 2-3 wafers. 10 identical tools: 42 wafers vs. 150 traditionally, saving 72%.
Embedded Real-Time Optimization
Unlike traditional offline simulation tools or standalone APC software, NeuroBox is directly embedded in the equipment control system via SECS/GEM for real-time closed-loop control.
R2R Technical Specifications
| Optimization Algorithm | CVXPY + OSQP Constrained Convex Optimization |
| Solve Latency | < 50 ms per iteration |
| Safety Mechanisms | Shadow mode, auto-rollback, interlock verification |
| Optimization Modes | Per-wafer / Per-lot / Manual+auto hybrid |
核心技术三:EIP 设备智能诊断 Core Technology 3: EIP Equipment Intelligence Platform
EIP 是 NeuroBox 的设备智能诊断平台,集成多模态数据融合与 AI 分析,实现 95% 以上的故障诊断准确率。
多模态故障诊断
- Run sheet 分析
- 设备日志解析
- 量测数据关联
- 视觉分析(腔室、晶圆表面)
设备健康评分
- 0-100 量化评分
- 部件状态评估
- 性能退化趋势
- 历史故障频率
预测性维护 (PM)
- 零部件寿命预测
- 维护排程优化
- 备品备件预测
- 停机时间最小化
产能分析 (WPH)
- 瓶颈识别
- 效率优化建议
- 产能预测
- 利用率提升
可检测问题类型:颗粒/污染、温度/压力异常、倾斜/对准偏差、膜厚/均匀性问题、RF 功率异常、气体流量异常等。
EIP is NeuroBox's equipment intelligence platform, integrating multi-modal data fusion and AI analysis to achieve 95%+ fault diagnosis accuracy.
Multi-Modal Fault Diagnosis
- Run sheet analysis
- Tool log parsing
- Measurement data correlation
- Vision analysis (chamber, wafer surfaces)
Equipment Health Scoring
- Quantitative 0-100 score
- Component state assessment
- Performance degradation trends
- Historical fault frequency
Predictive Maintenance
- Parts life prediction
- Maintenance scheduling
- Spare parts forecasting
- Downtime minimization
WPH Analysis
- Bottleneck identification
- Efficiency optimization
- Capacity forecasting
- Utilization improvement
Detectable issues: particle/contamination, temperature/pressure anomalies, tilt/alignment deviation, film thickness/uniformity, RF power anomalies, gas flow irregularities, and more.
硬件规格 Hardware Specifications
工业级无风扇设计,专为半导体洁净室环境打造。
Industrial-grade fanless design, purpose-built for semiconductor cleanroom environments.
计算与内存 Compute & Memory
| CPU | 8-core ARM Cortex-A78AE @ 2.0 GHz |
| GPU | 1024-core NVIDIA Ampere (100 TOPS INT8) |
| RAM | 16 GB LPDDR5 (102.4 GB/s) |
| 存储 Storage | 256 GB NVMe SSD + M.2 扩展expansion (max 2TB) |
| 计算平台 Platform | NVIDIA Jetson Orin NX 16GB |
通信接口 Communication Interfaces
| 千兆以太网 Gigabit Ethernet | 2x (设备端 + 厂务网络Equipment + FAB network) |
| USB 3.2 | 2x Type-A (10 Gbps) |
| RS-232 | 2x DB9 (SECS-I 支持support) |
| RS-485 | 1x (921600 baud) |
| GPIO | 8 通道channels (3.3V) |
物理与环境 Physical & Environmental
| 尺寸 Dimensions | 200 x 150 x 55 mm |
| 重量 Weight | ~1.2 kg |
| 散热 Cooling | 全被动散热(无风扇,适合洁净室) Passive heatsink (fanless, cleanroom-safe) |
| 工作温度 Operating Temp | 0°C ~ +50°C |
| 功耗 Power | 12W (待机idle) ~ 45W (满载peak) |
| MTBF | > 50,000 小时hours |
| 安装方式 Mounting | DIN rail / VESA 100mm / 壁挂Wall bracket |
协议与标准 Protocols & Standards
| SEMI E4 | SECS-I (串口通信Serial) |
| SEMI E5 | SECS-II (消息编码Message encoding) |
| SEMI E30 | GEM (通用设备模型Generic Equipment Model) |
| SEMI E37 | HSMS (TCP/IP) |
应用场景 Application Scenarios
NeuroBox 内置 10+ 种半导体工艺的物理模型,开箱即用。
离子注入 (Implant)
SRIM 物理模型,预测方块电阻(Sheet Resistance)。已在 Axcelis Quantum 设备上验证。
等离子刻蚀 (Etch)
RIE 模型,预测刻蚀速率与深度。支持 Lam Kiyo 等设备。
物理气相沉积 (PVD)
溅射产额模型,预测薄膜厚度与均匀性。支持 AMAT Centura 等设备。
化学气相沉积 (CVD)
Arrhenius 动力学模型,预测沉积均匀性。
化学机械抛光 (CMP)
Preston 方程模型,预测去除速率与均匀性。支持反向 recipe 推荐。
原子层沉积 (ALD)
ALD 生长模型,精确控制纳米级薄膜沉积。
NeuroBox includes built-in physics models for 10+ semiconductor processes, ready to deploy out of the box.
Ion Implantation
SRIM physics model for Sheet Resistance prediction. Validated on Axcelis Quantum equipment.
Plasma Etch
RIE model for etch rate and depth prediction. Supports Lam Kiyo and more.
PVD / Sputtering
Sputtering yield model for film thickness and uniformity. Supports AMAT Centura.
CVD
Arrhenius kinetics model for deposition uniformity prediction.
CMP
Preston equation model for removal rate and uniformity. Supports inverse recipe optimization.
ALD
ALD growth model for precise nanometer-level film deposition control.
与业界方案对比 Industry Comparison
与行业主流 APC / TCAD 仿真方案对比:
| 能力 | 传统 APC 方案 | 离线 TCAD 仿真 | NeuroBox™ E5200 |
|---|---|---|---|
| 物理模型 | 简单 R2R 控制 | TCAD 仿真 | 物理方程 + 贝叶斯标定 |
| 实验设计 | 无 / 人工经验 | 固定 DOE | 智能 DOE + 主动学习 |
| 需要晶圆数 | 25-50 片 | 30+ 片 | 10-15 片 |
| 反向 recipe 推荐 | 无 | 无 | 有(spec → recipe) |
| 机台间迁移 | 无 | 无 | 有(先验迁移,2-3 片) |
| 过程窗口分析 | 无 | 有(离线) | 有(实时,嵌入设备) |
| Cpk / spec 概率 | 无 | 无 | 有(UQ 量化输出) |
| 集成方式 | 独立软件系统 | 离线仿真工具 | 嵌入 SECS/GEM 设备控制 |
| 部署方式 | 服务器 / 云端 | 服务器 / 云端 | 边缘盒子(数据不出厂) |
Compared with mainstream APC and TCAD simulation solutions:
| Capability | Traditional APC | Offline TCAD Simulation | NeuroBox™ E5200 |
|---|---|---|---|
| Physics Model | Simple R2R Control | TCAD Simulation | Physics Eq. + Bayesian Calibration |
| Experiment Design | None / Manual | Fixed DOE | Smart DOE + Active Learning |
| Wafers Required | 25-50 | 30+ | 10-15 |
| Inverse Recipe | No | No | Yes (spec → recipe) |
| Tool-to-Tool Transfer | No | No | Yes (prior transfer, 2-3 wafers) |
| Process Window | No | Yes (offline) | Yes (real-time, embedded) |
| Cpk / Spec Probability | No | No | Yes (UQ output) |
| Integration | Standalone Software | Offline Simulation Tool | Embedded SECS/GEM |
| Deployment | Server / Cloud | Server / Cloud | Edge box (data stays on-site) |
获取更多信息 Get More Information
了解 NeuroBox™ 如何为您的设备赋予 AI 智能
Discover how NeuroBox™ can bring AI intelligence to your equipment
扫码关注公众号,获取最新产品动态 Follow our WeChat for the latest updates