A Taipei semiconductor supplier lifts wafer yield 5.4 points with engineer-annotated excursion detection

Context

Your Taipei semiconductor supplier ships specialty wafers to four of the world top-ten foundries. Yield on the highest-margin product line averaged 71%, against an industry benchmark of 78%. Each percentage point of yield gap represented approximately TWD 320M (US$10M) per year in lost contribution margin. The fab generated 14 TB of process telemetry per day across 1,800 sensors, but the engineering team spent 60% of their analytical capacity assembling pivot tables in Excel rather than diagnosing yield excursions. The CTO knew the data held the answer. The team did not have the tools or the time.

Challenge

Three constraints. First, the existing data warehouse predated the high-throughput sensors and choked on more than 24 hours of telemetry queries. Second, process engineers held tribal knowledge that no documentation captured: which sensor noise patterns were benign, which signalled an excursion, which were artefacts of a specific recipe. Third, fab engineers had been told four times in five years that "the data lake will fix this" by IT teams who then failed to deliver.

Approach

We ran a 4-phase model: foundation, surface, predict, hand-over. Foundation (8 weeks) replaced the bottleneck warehouse with a columnar time-series store, ingested 6 months of historical telemetry, and ran a 50-point validation against engineer-known events. The validation was the trust-building step: if the system did not match what engineers already knew, nothing else would matter.

Surface (6 weeks) built an ops dashboard for fab engineers, co-designed across 11 design sessions. The dashboard surfaced excursion candidates ranked by likely yield impact, with drill-through to the underlying sensor traces and a comment thread for engineer annotation. Engineer annotations fed back into the model as labelled training data.

Predict (10 weeks) deployed a yield-excursion early-warning model trained on 18 months of historical incidents and the engineer annotations from the surface phase. The model produced predictions 6-14 hours ahead of the typical detection window. Hand-over (parallel from week 18) trained two of your fab data engineers and the lead process engineer on model retraining, dashboard configuration, and the annotation workflow.

Results

Yield on the target product line rose from 71% to 76.4% over six months, a 5.4-point gain worth approximately TWD 1.7B (US$54M) in annualised contribution margin. Mean time to detect an excursion fell from 9.4 hours to 1.6 hours, an 83% reduction. Time spent assembling pivot tables fell from 60% of engineering capacity to 18%, freeing roughly 14,000 engineer-hours per year for actual diagnosis work. Three excursions in the first quarter were caught by the early-warning model before any wafer was lost; the engineering team estimated those events would have cost TWD 240M (US$7.6M) under the prior detection regime.

Two of the four foundry customers cited the yield improvement in their quarterly supplier reviews, securing volume commitments for the next product cycle.

Lessons

The 50-point validation against engineer-known events was the trust gate that the prior four IT-led attempts had skipped. Engineer annotation as a feedback loop turned the dashboard from a one-way reporting tool into a system that learned from its users. Replacing the warehouse first, before any modelling, was the unglamorous step that made everything else possible.