Case File 02 · Semiconductor
Lifting yield with inline defect analysis
Representative — advanced-node fab
▸ ENGAGEMENT DETAILS
Representative scenario modeled on published industry benchmarks. Not a named client.
The Challenge
A representative advanced-node fab needs to find the defect — and its root cause — before the lot ships.
Manual and rule-based visual inspection struggles with micro-cracks and sub-surface anomalies, and a defect that escapes inspection is only caught later, when the cost of the lot is already sunk. McKinsey puts yield loss and test at 20-30% of total production cost, and estimates that machine-learning root-cause analysis can reduce yield detraction by up to about 30%.
The harder problem is scale, not accuracy in a lab. Survey work suggests only around 17% of ML vision inspection has reached high-volume production; the rest is still in pilot. Closing that gap — integration, evaluation, and governance — is what this engagement is scoped around.
Our Agent Solution
We implement ChipSense — our semiconductor agent system — against the fab's own inspection and metrology data. Models are tuned and evaluated on the client's historical defect library under a signed data agreement and validated against expert inspector review. Inspection agents classify defects by type and severity across magnification levels; an analysis agent correlates defect signatures with process and FDC parameters to propose a likely root cause. Every process adjustment leaves the system as a recommendation routed to a process engineer — never as an automatic change to a tool.
Inspection Agent
Analyzes wafer images at multiple magnification levels and classifies candidate defects
Analysis Agent
Classifies defects by type and severity, then correlates signatures with process and FDC parameters
Optimization Agent
Proposes process adjustments for engineer review, with the supporting correlation attached
Reporting Agent
Generates quality reports, drift alerts, and trend analysis for the engineering team
Implementation Timeline
A representative 16 weeks delivery path, from discovery to handover.
Data Audit & Assessment
Audited the existing defect library and its labeling quality, assessed imaging and metrology equipment, mapped the inspection workflow
Weeks 1-3
Model & Agent Architecture
Designed the multi-resolution inspection pipeline, defined the defect taxonomy with the fab's engineers, set the human-review model
Weeks 4-6
Build & Evaluation
Built the agent system, tuned and evaluated it on the client's own historical corpus under a signed data agreement, validated against expert inspectors
Weeks 7-12
Integration & Calibration
Integrated with production-line imaging and MES, calibrated thresholds per product line, ran alongside the existing inspection process
Weeks 13-15
Production Handover
Phased cutover with monitoring, drift alerts, an operator runbook, and a documented rollback path
Weeks 16
Target ranges — and today's baseline
Both sets of figures are drawn from published research, not from VelocityMind client results. Targets are the ranges this scenario is scoped against; baselines describe where the status quo sits. Actual results depend on your data, systems, and scope.
Less yield detraction
Source · McKinsey — ML root-cause analysis, up to ~30%
30%
Defect-classification accuracy (ceiling)
Source · Peer-reviewed deep-learning machine vision, >97-99%
99%
Lower manufacturing cost at scale
Source · McKinsey, 2021 — AI at scale in semiconductor
17%
Where the baseline sits today
Published measurements of the problem, shown so the target ranges above can be read against something. These are not outcomes we delivered.
Of production cost is yield loss and test today
Source · McKinsey — 20-30% of total production cost
25%
What ships — and what stays human
The artifacts you own at handover, and the decisions the agents never take.
▸ DELIVERABLES
What you receive
- Inspection and analysis agents running against your own imaging and metrology feeds
- Defect taxonomy and labeling standard agreed with your engineering team
- Evaluation harness with per-product-line thresholds and a documented accuracy baseline
- MES and imaging integration contracts, plus drift alerts and a rollback path
- All source code in your repository, with the deployment owned by your team
▸ GUARDRAILS
What stays human
- Process adjustments are recommendations. A process engineer approves every change; the system never writes to a tool.
- The system runs alongside the existing inspection process through calibration, so the fab keeps a comparison baseline.
- Accuracy figures are stated as a ceiling from published studies, not a commitment — real performance depends on your defect library and imaging.
- Proprietary defect data stays inside the fab's environment under the signed data agreement and is not used to train shared models.
Method and sources
This scenario is constructed, not reported. Every figure on this page traces to one of the published sources below.
ML root-cause analysis reduces yield detraction by up to ~30%
McKinsey
Yield loss and test account for 20-30% of total production cost
McKinsey
AI at scale can cut semiconductor manufacturing cost by up to ~17%
McKinsey, 2021
Deep-learning machine vision reaches >97-99% wafer/defect classification accuracy
Peer-reviewed studies
Only ~17% of ML vision inspection is in high-volume production — the rest is pilot
ScienceDirect survey
Ranges are published industry benchmarks. They describe what programs of this shape have achieved elsewhere, not a commitment for your environment — the range we would scope against for you is set after discovery.
A representative perspective
“Inline inspection and root-cause analysis surfaced patterns our manual process missed, and the suggestions moved yield in the right direction.”
▸ NEXT STEP
Run this against your own workflow
Send us one workflow, its monthly volume, and the systems it touches. We reply within one business day with a first read on whether an agent system is the right tool, the agent shape we would propose, and an engagement range.
No commitment. We reply within one business day.