Data Sources for Yield

The Yield Data Landscape

Yield prediction requires integrating multiple data sources across the entire manufacturing flow:

• Inline metrology: CD, overlay, film thickness, and other measurements taken during fabrication. Sparse sampling (5–20 sites per wafer, 5–10% of wafers).
• FDC (equipment sensor data): Process conditions for every wafer on every tool. Complete coverage but indirect — must be correlated to yield outcomes.
• Defect inspection: Defect counts, maps, and classifications from optical and e-beam inspection.
• WAT (Wafer Acceptance Test): Electrical measurements on test structures after fab completion — transistor parameters (Vt, Idsat, Ioff), resistances, capacitances.
• Sort/probe data: Die-level pass/fail and bin results from electrical testing.
• Design data: Die layout features — pattern density, metal coverage, critical design rules.

Building the Wafer × Die Feature Table

Every yield model in production starts from one canonical artifact: a flat (wafer, die) feature table. Building it cleanly determines whether the model works.

Standard schema

Sketch of the build pipeline

import pandas as pd

def build_die_feature_table(lot_ids):
    """Join MES + FDC + metrology + defects + WAT + sort into a die-level table."""
    mes      = load_mes_history(lot_ids)              # wafer × step
    fdc      = load_fdc_summaries(lot_ids)            # wafer × step
    metro    = load_inline_metrology(lot_ids)         # site
    defects  = load_defect_records(lot_ids)           # die
    wat      = load_wat(lot_ids)                      # site
    sort     = load_sort_bins(lot_ids)                # die

    # 1. Wafer-level: route + FDC summaries
    wafer_df = mes.merge(fdc, on=["lot_id", "wafer_id", "step"])
    wafer_df = wafer_df.pivot_table(
        index=["lot_id", "wafer_id"],
        columns="step",
        values=[c for c in wafer_df.columns if c.startswith("fdc_")],
    )
    wafer_df.columns = ["__".join(c) for c in wafer_df.columns]
    wafer_df = wafer_df.reset_index()

    # 2. Site-level → die-level by nearest-neighbor on (x, y)
    metro_die  = interpolate_to_dies(metro,  key=("die_x", "die_y"))
    wat_die    = interpolate_to_dies(wat,    key=("die_x", "die_y"))

    # 3. Defect counts per die (0.5 mm radius)
    defect_die = count_defects_per_die(defects, radius_mm=0.5)

    # 4. Final outer join
    die_df = sort.merge(wafer_df,  on=["lot_id", "wafer_id"], how="left")
    die_df = die_df.merge(metro_die, on=["lot_id", "wafer_id", "die_x", "die_y"], how="left")
    die_df = die_df.merge(wat_die,   on=["lot_id", "wafer_id", "die_x", "die_y"], how="left")
    die_df = die_df.merge(defect_die,on=["lot_id", "wafer_id", "die_x", "die_y"], how="left")
    return die_df

Three things that go wrong

• Step-name drift — recipes get renamed; you suddenly have two columns for the same physical step
• Coordinate misalignment — inline metrology sites and probe dies use different coordinate systems; mis-mapping invents fake correlations
• Look-ahead leakage — joining WAT (taken after the fab) into a feature for predicting yield is fine; joining it as a process input is data leakage