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.
Four stages from raw inspection data to actionable yield intelligence.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
Top-1 accuracy on CDSEM review images, 14nm and below nodes, 12 defect type families
Per wafer map, on-premise GPU hardware (NVIDIA A30 or equivalent). No cloud round-trip.
Median time from inspection event completion to engineer notification with full evidence package
Pre-trained defect type library covering particle, pattern, CMP, etch, lithography, and deposition families
SEMI E5/E30, KLA KLARF, ASML IBIS, Applied Materials ExL format, Lam Insights API
100% on-premise or private cloud. No outbound data transfer required for model inference or alerting.
SynthKernel connects to the equipment and MES systems already running in your fab. No rip-and-replace.
We configure a proof-of-concept using your existing inspection output — no new tooling required — and deliver a yield impact assessment within four weeks.