+2.4pp yield uplift across three fabs, £10m+ annualised.
A semiconductor IDM ran inline metrology, probe and final-test data in three disconnected stacks. A unified yield-analytics platform with ML-based root-cause tooling delivered 2.4pp yield uplift on the target node in eight months.
From siloed fabs to shared yield loop.
Each fab had its own DW, its own yield-engineering team, and no cross-fab comparability. Root-cause work took weeks. The target was a shared platform that could surface cross-fab patterns in hours, and an ML model surface to accelerate the engineer's intuition, not replace it.
Approach.
Platform foundation
Databricks on AWS, open data model across inline, probe, final-test, equipment.
Unified ingestion
CDC from fab systems, schema alignment, lot-to-device traceability preserved.
Yield ML surface
Yield-prediction models per node, Shapley attribution for engineers.
Engineering UX
Tool built for yield engineers, not data scientists — reviewed weekly with them.
Rollout
Fab-by-fab, preserving each site's existing reviews while converging analytics.