Deployment & Operations

Deploying PdM in Production

Moving from a Jupyter notebook to a production PdM system involves significant engineering:

• Data pipeline: Real-time ingestion of FDC data from 1,000+ tools, cleaning, feature computation, and storage. Must handle missing data, sensor failures, and recipe changes.
• Model serving: Low-latency inference after each process run (seconds, not minutes). Models must handle multi-chamber, multi-recipe scenarios.
• Alert management: Converting model scores into actionable alerts. Too many false alarms = alert fatigue (engineers ignore them). Too few = missed failures.
• Integration with MES: Alerts flow into the Manufacturing Execution System for maintenance scheduling and wafer routing decisions.
• Model monitoring: Track model performance over time. Equipment changes (new PMs, recipe updates) can invalidate models — requiring retraining or adaptation.

Alert Thresholds, Drift, and the MES Feedback Loop

A PdM model is only as good as the decisions it triggers. Three operational pieces dictate whether the savings actually land.

1. Setting alert thresholds

Most fabs adopt a tiered alert scheme — typically a Yellow / Orange / Red triage:

2. Model drift

Equipment evolves: new PMs, new chambers, recipe edits, target swaps. A model trained six months ago can quietly become useless. Monitor drift continuously:

from scipy import stats

def feature_drift(train_dist, recent_dist, alpha=0.01):
    """Return True if the recent feature distribution has drifted (KS test)."""
    ks_stat, p_value = stats.ks_2samp(train_dist, recent_dist)
    return p_value < alpha, ks_stat

# Concept-drift retraining trigger
drifted, score = feature_drift(
    train_dist=feature_history["chamber_pressure_mean"][:30_000],
    recent_dist=feature_history["chamber_pressure_mean"][-2_000:],
)
if drifted:
    schedule_retrain(model_id="etch_chamber_rul", reason=f"KS={score:.3f}")

3. The MES loop

The output of the PdM system is not a CSV — it is a structured event posted to the Manufacturing Execution System (MES):

• A predicted failure event creates a maintenance ticket in the CMMS
• The scheduler reserves the tool for PM at a low-WIP window
• Wafer routing is rebalanced to peer chambers
• Once PM is closed, the model receives a labelled failure or no-failure event for future retraining