Large Universe Models in semiconductor manufacturing

What it ingests

In-line metrology, equipment sensor telemetry, defect inspection images, wafer electrical test, maintenance logs and chamber histories, materials lot genealogy, and environmental data.

The belief that matters

Attribution of yield loss to process steps, chambers and material lots, maintained continuously instead of reconstructed after each excursion.

Excursion detection versus excursion explanation

Statistical process control detects that something moved. It does not say what. Explanation is a retrospective investigation that takes days, during which the fab keeps producing.

A Large Universe Model that maintains attribution beliefs continuously has already accumulated the evidence when the excursion occurs — which chamber drifted, which lot genealogy is over-represented, which maintenance event preceded the shift.

Correlated evidence

The hard cases are interactions: a chamber that is fine except with a particular material lot at a particular ambient humidity. Holding these as joint beliefs updated by every wafer is tractable in a way that per-variable control charts are not.

Limits

Fabs already run extensive statistical process control and fault detection built by people who know the physics. A general model contributes breadth — pulling maintenance text, supplier notices and environmental data into the same belief — rather than replacing the specialised instruments.