Rethinking Cycle Time Drivers in High-Mix Semiconductor Fabs

To compare theory against real‑world behavior, 1.5 years of detailed factory data were collected. The dataset included:

• Equipment availability
• Process time variability
• Utilization
• Number of process tools per area

Using the CRISP‑DM methodology, the data was carefully cleaned to remove events that could distort cycle time results. This includes shutdown periods, staging steps, low‑volume equipment types, and toolsets with weak integration.

After cleaning, approximately 12% of the original records were removed, leaving roughly 2,200 weekly observations aggregated by equipment type.

This level of longitudinal and granular fab data has historically been difficult to obtain. It enabled a deeper examination of which factors truly influence cycle time variation in practice.

Even after introducing additional derived factors to improve model accuracy, the analysis showed that some important influences remain unmodeled and continue to affect fab performance.

The results highlight several important findings.

The findings show that process time variability has the strongest overall influence on cycle time, with a Pearson correlation of 0.13. However, this relationship looks very different when broken down by equipment type, from strongly positive (0.83) to moderately negative (‑0.24). These negative correlations point to missing factors. Because many of these tools are batch systems, changes in product mix, and therefore recipe diversity, are likely driving unexpected shifts in runtime and utilization.

The analysis also showed that process time itself has a negative correlation with cycle time (‑0.13). Since wait time includes a fixed transport component, its influence becomes relatively smaller as process time increases.

Equipment availability (‑0.10) and utilization (‑0.10) also showed weaker effects than predicted by earlier sensitivity analysis. In non‑focus toolsets, utilization even correlated negatively with cycle time. As these tools become busier, they tend to receive more support such as additional labor, which offsets the wait‑time impact that theory would typically predict.

The study identified several improvements that would enhance future cycle time modeling efforts:

• Tracking transport time directly or subtracting it from cycle time
• Measuring batch tool behavior, such as average batch utilization and recipe diversity
• Introducing a manual classification of focus vs. non‑focus toolsets

These additions would help capture influential factors currently missing from existing models.