Classification & Root Cause

ML for Root Cause Analysis

Beyond detecting defects, ML helps identify what caused them:

• Equipment fingerprinting: Each process chamber leaves subtle "signatures" on wafers. ML models can identify which specific chamber processed a wafer based on defect patterns or metrology signatures — essential for isolating problematic equipment.
• Correlation analysis: Linking defect occurrences to upstream process parameters. Random Forest feature importance or SHAP values reveal which equipment parameters most strongly predict defects.
• Temporal analysis: Tracking defect rate trends after PM events, recipe changes, or chemical lot changes to identify root causes.
• Spatial signature matching: Comparing wafer-level defect patterns against a library of known signatures. Each root cause (reticle defect, chuck contamination, edge ring wear) produces a characteristic spatial pattern.

Commonality Analysis and SHAP-Based Diagnosis

The first question an FA engineer asks when yield drops is "what do all the bad wafers have in common?" — and the second is "which sensor moved?" Two ML-friendly techniques cover both.

1. Commonality analysis

For each candidate factor (chamber, recipe, photoresist lot, operator shift), compute the defect rate among wafers that touched that factor vs. those that didn't. A Fisher exact test or χ² ranks the most likely culprit:

import pandas as pd
from scipy.stats import fisher_exact

def commonality_rank(df: pd.DataFrame, defect_col="is_killer",
                     factor_cols=("etch_chamber", "litho_chamber", "resist_lot", "shift")):
    """Rank factors by association strength with defect occurrence (Fisher p-value)."""
    rows = []
    total_bad = df[defect_col].sum()
    total = len(df)
    for f in factor_cols:
        for value, group in df.groupby(f):
            a = group[defect_col].sum()
            b = len(group) - a
            c = total_bad - a
            d = (total - len(group)) - c
            _, p = fisher_exact([[a, b], [c, d]], alternative="greater")
            rows.append({"factor": f, "value": value, "p_value": p,
                         "bad": int(a), "total": len(group)})
    return pd.DataFrame(rows).sort_values("p_value").head(20)

2. SHAP for parametric root cause

When the suspect is a sensor (not a categorical factor), train a gradient-boosted classifier that predicts pass/fail from upstream FDC summary stats and inspect SHAP values:

import shap
from xgboost import XGBClassifier

model = XGBClassifier(n_estimators=300, max_depth=4, learning_rate=0.05)
model.fit(X_train, y_train)

# Per-wafer attribution: which features push prediction toward "fail"?
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(X_test)

# Top global drivers across the lot
import numpy as np
mean_abs_shap = np.abs(shap_values).mean(axis=0)
top10 = np.argsort(mean_abs_shap)[-10:]
for i in reversed(top10):
    print(f"{X_test.columns[i]:30s}  |SHAP|={mean_abs_shap[i]:.4f}")

Layering the two

Commonality analysis narrows the search to a chamber or a lot; SHAP on that subset pinpoints which sensor signature is doing the damage. Together they cut root-cause-investigation time from days to hours.