Align With Existing Process Flows & Tool Architecture, Converge New Tools to Match Golden Baselines
When qualifying a new tool or chamber, the core requirement is matching existing process flows, tool architectures, and Golden reference baselines. The system aligns wafer spatial distributions, key metrology metrics, and hardware/Recipe configurations against Golden references, quantifying mismatches and suggesting grounded tuning directions.
From Excursion Detection to Handoff & Disposition, Rapidly Build Comprehensive Evidence & Boundary Guidance
Real-world defect triage is far more complex than subtracting two wafer maps. From initial Excursion detection, the system rapidly gathers and aligns spatial maps, SEM morphology, cross-step routing histories, and FDC traces to define affected regions and candidate causes—forming an auditable evidence pack for smooth cross-team Handoff and lot Disposition.
03 · ON-TOOL ML MODEL DRIFT & IN-LINE HYPERPARAMETER TUNING
Address On-Tool CNN & Regression Model Drift, Provide In-Line Hyperparameter Tuning Guidance
On-tool ML models—typically combining CNN defect classification models with Correlation Regression models—often drift when process conditions or film stacks shift. Without exporting sensitive data, the system diagnoses CNN false alarms and regression residuals in-line, recommending hyperparameter, weight, and threshold updates verified on local datasets.
CNN + Regression TuningDiagnose on-tool ML drift · In-line hyperparameter guidance · Local backtesting
Extract Experience from 8D Reports, PDFs & Slides, Solve Lost Know-How Across Turnover on ~80% Recurring Issues
Roughly 80% of equipment bring-up and line excursion issues in a Fab have occurred before, yet solutions sit scattered across historical 8D reports, handover Slides, and technical PDFs—often lost when engineers rotate. The system extracts and retains this knowledge so teams can immediately reference proven paths instead of repeating old detours.
Let Agent operate the workspace. Keep every step visible to engineers.
The Agent inspects imagery and prior runs, executes standard checks first, and runs targeted analyses on demand.Grounding decisions in accumulated evidence and verified engineering method specs.
Investigation walkthrough playing · Switch steps anytime
Turn raw data, into grounded engineering recommendations.
Automates routine analysis across more wafers, time windows, and operating conditions, surfacing valuable clues and candidate causes. Engineers inspect evidence layer-by-layer, steer investigation directions, and advance findings into tuning and validation plans.
Diagnostic reports and recommendations, backed by complete investigative evidence.
A comprehensive report outlining current diagnosis, recommended actions, and validation procedures, with every suggestion tied to analytical traces, evidence, and open questions.
Tool & Flow Matching
Excursion to Disposition
CNN / Regression Tuning
Report Deliverable Overview
New Tool & Golden MatchingEE / PE · Equipment & Process Engineering
New-Tool Golden Matching & Architecture Alignment Assessment
Diagnostic Report
Spatial deviations vs. Golden baseline, process flow compatibility, and tool architecture (chamber/thermal/Recipe) matching; candidate factors and evidence.
Recommendations
Architecture-compatible priority checks, tuning sequence, and grounded parameter adjustment directions.
Validation Plan
Defined Golden reference baselines, re-measurement conditions, and pass/fail criteria for matching, stability, and flow compatibility.
Full Evidence Traces
Target and Golden reference datasets, hardware architecture and Recipe snapshots, computation logs, and uneliminated candidate factors.
Evidence chain connecting findings to raw data and execution traces
Extract know-how from 8D, PDFs & Slides, into retained Agent method specs.
Extracted from historical 8D reports, troubleshooting PDFs, handover Slides, and process specs, bridging Fab engineer turnover gaps to address roughly 80% of historically recurring line issues.
EXPERTISE / 8D · PDF · SLIDES RETENTIONMethod Codification Flow
01 Field Experience
02 Expert Review
03 Spec Execution
04 Continuous Asset
“Verify scan area first, compare defect distributions second.
Awaiting expert-reviewed specs and active task integration.
↺New Cases & Field Engineering DocsExtract Candidate Know-How → Expert Review → Spec Release → Prevent Turnover Loss
SAME HARNESS / NEW SPECS & TOOLS
One architecture, scalable to new engineering domains.
Method specs govern how judgments are formed; tool contracts define what actions can be executed. New engineering tasks simply plug in validated tools.
Excursion to Disposition TriageActive Production Pilot
CNN / Regression In-Line TuningExpansion Track
Tool Matching & Recipe TuningExpansion Track
Wafer Analysis: Comparability specs, scientific computing tools, and evidence standards together power continuous investigation.
Experience belongs to the team, surviving staff turnover.
Extracts and retains troubleshooting know-how from historical 8D reports, PDFs, and handover Slides—covering ~80% of recurring Fab issues.
Specs guide every step of the workflow.
The Harness bridges specifications with ingested lot data, compute tools, and runtime verification; the Agent operates strictly within valid bounds.
Capabilities scale cleanly through specs.
Expand to new engineering tasks with modular spec packages, tool contracts, and benchmark lots while sharing the proven runtime harness.
Engineering expertise is managed as modular specs, published after expert review, and reused across future investigations.
ArchitectureHow Experience, Computing, and Evidence CollaborateView Architecture
PURPOSE-BUILT AGENT ARCHITECTURE
Designed specifically for semiconductor image analysis.
TranslatingFab engineering expertise and senior analytical methodsinto applicability prerequisites, scientific computing routines, and rigorous evidence standards. Our Agent architecture advances along these engineering constraints to form a closed-loop investigation.
FAB ENGINEERING EXPERTISE
Understanding the cleanroom: knowing what questions to ask.
Incorporates tool state, recipe parameters, review optics, and comparison goals into the run, pinpointing which clues warrant DOE and what data must be fetched.
SENIOR ANALYTICAL METHODS
Definitive methods: knowing how to validate hypotheses.
Verifies population sample sizes, active inspected areas, analysis metrics, and baseline definitions to clearly distinguish observations from engineering explanations.
AUTONOMOUS AGENTInvestigation & PlanningInspects available data, selects next action, and steers trajectory based on quantitative metrics.
↔
SCIENTIFIC EXECUTIONAnalysis & Scientific ComputeExecutes rigorous algorithms on defined inputs, returning fully auditable calculations.
→
EVIDENCE & PROVENANCEFindings & Evidence TracesLinks raw wafer imagery, computational formulas, result revisions, and remaining evidence gaps.
Observe→Select Action→Execute→Verify→Update
Engineer · Confirm scope, pause, intervene, review findings, and schedule DOE
How do I audit its conclusions?
07 / EVIDENCE & AUDITABILITY
White-box AI Agent: every recommendation backed by verifiable evidence.
From full-wafer imagery and recipe parameters to actionable engineering recommendations. Inputs, method specs, computations, and execution traces are auditable layer by layer.
Dual-Source InputsWhite-Box Analytics · Every Step AuditableEngineering Recommendations
Data & Compute Relations Hypotheses Pending Validation
Engineering AdvisoryAudit Variations First, Isolate Causes SecondInspection & Unit Steps · PE / PIEScreen candidate hypotheses before scheduling validation DOE
FINDINGOuter Ring Defect ConcentrationImage Review · Finding F-02
VALIDATIONCandidate Hypotheses Pending ValidationEvidence Gap Logged · Next Check Scheduled
From inputs to recommendations, every layer of evidence opens automatically.
02 / 04
Every investigative branch leads directly to its underlying evidence.
WHITE BOXEvidence Retained with Findings
Traceable InputsImage and recipe context snapshots
Grounded MethodsDomain expertise and verified applicability bounds
Verifiable ComputeStandardized formulas and statistical parameters
Logged ExecutionTool call parameters and result revisions
Engineering Advisory
Confirm temporal deltas before isolating root causes.
Audit inspection conditions and intra-wafer progression first, then leverage recipe and tool logs to isolate candidate hypotheses. Recommended parameter adjustments require direct empirical evidence and validation DOEs.
Advisory Bounds & ConfidenceSpatial Distribution Quantified · Root Cause Pending Validation
Retains supporting evidence, gaps, and non-excluded candidate hypotheses. Recommendations provided for engineer review; correlated leads still require validation.
What engineering experience do we want to deliver?
08 / OUR MISSION
Exploring the next generation of semiconductor analysis and validation.
Starting from Fab image analysis and yield learning, we bring engineers and specialized AI agents onto a unified workbench to investigate continuously, compute rigorously, and verify every finding together.
Transforming one-off reports into an ongoing engineering journey. New problems, fresh lot data, and emerging discoveries connect seamlessly inside the same shared context.
RESOURCES & DOCUMENTATION
Explore use cases, and inspect our engineering methods.
