From inspection data to corrective action — not days, minutes

SynthKernel processes inline inspection output in real time, correlates defect signatures across layers, and delivers ranked root-cause hypotheses before the next lot enters the affected process step.

How SynthKernel Works

Four stages from raw inspection data to actionable yield intelligence.

01

Data Ingestion

Connects to KLA, ASML, Lam, Applied Materials, and TEL tools via SECS/GEM and SEMI E10/E30 interfaces. Supports KLA KLARF, ASML IBIS, and custom recipe export formats. No middleware installation required on tool controllers.

02

Defect Classification

CNN-based classifier processes CDSEM review images and optical scanner thumbnails. Bins are trained per technology node and defect type family — particle, pattern, scratch, CMP residue, bridging. Classifier confidence scores feed directly into the correlation engine.

03

Multi-Layer Correlation

Defect die maps from each inspection layer are cross-correlated with WAT and probe electrical test data. The correlation engine identifies which inspection-layer defect signature has the highest statistical association with failing die locations at final test.

04

Engineer Alert & Evidence Package

Yield engineers receive a prioritized alert with the estimated yield impact, a die-level overlay of defect versus fail maps, the list of lots on the suspect tool, and the specific recipe step flagged by the model. No additional investigation required to act.

Every major yield failure mode, covered

CDSEM and Optical Review Integration

Process review SEM images from KLA 29xx/39xx series and optical review images from KLA Surfscan and Puma platforms. The same classification pipeline handles both modalities with per-node model variants. No manual image labeling required for standard defect types covered in our pre-trained library.

Spatial Pattern Recognition

Detect systematic yield-killing spatial signatures — edge ring, center spot, arc, and scratch — directly from inline wafer inspection density maps. Spatial pattern detection runs before defect classification, enabling faster excursion identification when defect counts are low but distribution is non-random.

Process History Lineage

Each wafer event is linked to the exact equipment ID, chamber number, recipe version, and operator shift. When a defect cluster is identified, SynthKernel queries the lineage database to find the common process step across all affected wafers — even when they came from different lots entering the fab days apart.

Adaptive SPC and APC Integration

Feeds statistically significant defect density shifts back into existing SPC charts and APC feedback loops. When SynthKernel confidence in a process drift exceeds the configured threshold, it can trigger an APC correction on the affected tool directly — subject to your process change management policy.

Lot Disposition Workflow

When an excursion is confirmed, SynthKernel generates a disposition recommendation for all in-flight lots processed on the suspect equipment, including estimated die yield for each lot. Disposition decisions are recorded with supporting evidence for later review and audit.

Air-Gapped Deployment Option

All inference runs locally. Wafer images, die maps, process parameters, and equipment logs never leave the facility. The platform is validated for operation in ITAR-controlled environments and supports customer-managed encryption keys for all stored inspection data.

Built to production-fab tolerances

Classification Accuracy

97.3%

Top-1 accuracy on CDSEM review images, 14nm and below nodes, 12 defect type families

Inference Latency

43ms

Per wafer map, on-premise GPU hardware (NVIDIA A30 or equivalent). No cloud round-trip.

Alert Response Time

12 min

Median time from inspection event completion to engineer notification with full evidence package

Defect Type Coverage

47 types

Pre-trained defect type library covering particle, pattern, CMP, etch, lithography, and deposition families

Supported Tool Interfaces

SECS/GEM

SEMI E5/E30, KLA KLARF, ASML IBIS, Applied Materials ExL format, Lam Insights API

Data Residency

On-Premise

100% on-premise or private cloud. No outbound data transfer required for model inference or alerting.

Works with your installed toolset

SynthKernel connects to the equipment and MES systems already running in your fab. No rip-and-replace.

Inspection Equipment

KLA Surfscan SP7 KLA 2930 Review SEM ASML HMI eScan Lam Seguro Applied Materials Enlight Hitachi SU9300 Onto Innovation Iris

MES and Process Control

Workstream MES Camstar SEM IQT YieldStar GlobalFoundries SPC Synopsys Yield Explorer PDF Solutions Exensio

See it on your data, in your fab environment

We configure a proof-of-concept using your existing inspection output — no new tooling required — and deliver a yield impact assessment within four weeks.