Problem solving
Root Cause Analysis: From Symptom to Verified Cause
A copper-pillar example for separating symptoms, correlations, and causal evidence.
Quick answer
Root cause analysis is not finished when the team agrees on the most plausible explanation. It is finished when a cause has been shown, with evidence, to produce the problem, and removing that cause makes the problem stop.
Three things get confused along the way:
| Term | What it is | Example |
|---|---|---|
| Symptom | What was observed | Copper-pillar height measured at 42 µm against a 48–52 µm requirement. |
| Correlation | A condition that appears together with the symptom | Low readings only on lots processed on one tool. |
| Causal evidence | A demonstrated mechanism linking a condition to the symptom | Restoring a displaced component returns height to specification; displacing it again reproduces the shortfall. |
Most weak root cause reports stop at the second row and write it up as the third.
This guide follows the fictional copper-pillar case introduced in the PFMEA guide. All values are teaching values, not process recommendations.
Write a problem statement the data can test
How this case started
The pre-release height check on lot 2317 returned three wafers with edge readings between 41 and 43 µm. The lot was placed on hold, and the downstream joining step was told not to expect it.
The hold itself was containment. It stopped one lot from moving on; it did nothing about the next lot. Two more lots from tool B were in the queue, and nobody could yet say whether they would show the same result. The PFMEA for this step had listed several causes for low height, and none of the quick checks pointed to any of them: the recipe revision was correct, the bath had been sampled that morning, and the rectifier log showed the setpoint current.
That gap is the reason to start a root cause analysis. A problem that can be explained by a known cause gets a known fix. A problem that cannot be explained yet needs an investigation before any fix can be trusted.
A problem statement describes what was observed, not why. Keep the explanation out of it.
| Element | Statement |
|---|---|
| What | Copper-pillar height below the 48 µm lower limit. |
| Where | Lot 2317, tool B; measured sites near the wafer edge. |
| When | First detected at the pre-release height check, three days after tool B's scheduled maintenance. |
| How much | Edge sites 41–43 µm; center sites 49–51 µm. |
| How many | 3 of 25 wafers failed the sampled check; the other 22 passed. |
Compare that with "low height due to wrong recipe." The second version has already decided the cause before anyone looked at the data. If the recipe turns out to be correct, the problem statement is wrong, and the team has to start over.
Two details in the table matter later: the failure is localized (edge, not center), and it has a time marker (after maintenance). Neither is a conclusion yet. Both are things the data can test.
Stratify before you brainstorm
Before opening a fishbone diagram, split the existing data along the dimensions you already have. Each split either narrows the search or rules something out.
| Stratification | Fictional result | What it tells the team |
|---|---|---|
| By tool | Tools A and C in specification; tool B low. | Chemistry shared across tools is unlikely; tool-specific condition is likely. |
| By site within wafer | Center 49–51 µm; edge 41–43 µm. | Not a uniform thickness shift; points to current distribution, not total charge. |
| By time | Tool B lots before maintenance in specification; lots after maintenance low. | Something changed at or after the maintenance event. |
| By product | Same product on tool A is in specification. | Product design and photoresist opening are not the first suspects. |
| By recipe revision | Same recipe revision on all three tools. | Obsolete-recipe cause from the PFMEA does not explain this event. |
Five rows of stratification have done more than an hour of brainstorming would. They have taken the recipe cause off the table for this event, pushed bath chemistry down the list, and pointed at one tool during one time window.
Stratification depends on how the data were recorded. If height measurements were logged without site location, the center-versus-edge split is not available, and the team is working with less than it could. Trend charts help here: an I-MR chart of tool B edge heights would show whether the shift was a step change at the maintenance date or a gradual drift.
One check must come before any of this: confirm that the measurement system can resolve the differences you are about to interpret. A center-to-edge difference of 8 µm means little if the gauge's repeatability is 5 µm. If the gauge has not been characterized, run a Gage R&R first. Analyzing a measurement artifact as a process problem is a common way to lose two weeks.
Fishbone: organize candidates, don't rank them
The fishbone (Ishikawa) diagram organizes potential causes by category. It is a checklist for completeness, not a ranking tool and not a conclusion.
The PFMEA for this step listed several causes for low height: delivered current, plated area, bath composition, within-wafer uniformity, and photoresist geometry. That list is a good starting set of branches. The stratification results then do the first round of pruning: a uniform-charge cause (current setpoint, plated area) would lower all sites, not just the edge, so the Machine branch narrows to things that change current distribution across the wafer.
Common fishbone mistakes
| Mistake | What it looks like | Better practice |
|---|---|---|
| Treating the diagram as the answer | The report shows a fishbone and then states the root cause. | The fishbone lists candidates. Evidence selects among them. |
| Filling categories for symmetry | Every branch has exactly three items, including "operator attitude" under Man. | Leave a category sparse if nothing credible belongs there. |
| Writing symptoms as causes | "Low current density at edge" placed on a branch. | That is a mechanism, one step closer to the effect. Ask what would produce it. |
| Writing effects as causes | "Open joint at assembly" placed on a branch. | That is downstream of the effect, not upstream. |
| Skipping the Measurement branch | The gauge is assumed correct because it is calibrated. | Calibration addresses bias; it says nothing about repeatability or site-to-site reproducibility. |
| Brainstorming without the data | The fishbone is drawn in a meeting before anyone has stratified the results. | Stratify first so the diagram starts with half its branches already evaluated. |
| Drawing one fishbone for several problems | Edge low height and a separate void defect on the same diagram. | One effect per diagram. Different effects usually have different causes. |
5 Why: follow the mechanism, stop at something you can act on
The 5 Why technique asks why repeatedly until the chain reaches a condition that can be changed. The number five is a guideline; the stopping rule is what matters.
Each step in the chain needs its own evidence. A chain that is logically coherent but unsupported is a story, not an analysis.
| Why | Answer | Evidence that supports this step |
|---|---|---|
| Why are edge pillars low on tool B? | Current density at the wafer edge is lower than the recipe assumes. | Edge-only shortfall; center sites in specification; same recipe produces correct height on tools A and C. |
| Why is edge current density low? | The anode shield in tool B is positioned closer to the wafer than its setpoint, shadowing the edge. | Physical inspection: shield measured 6 mm from its marked setpoint. Tools A and C at setpoint. |
| Why is the shield out of position? | It was removed during the maintenance and reinstalled without being set to the reference mark. | Maintenance record shows shield removal; no entry for position check. |
| Why was the position not set? | The post-maintenance checklist has no step for shield position verification. | Checklist reviewed: contact ring, anode, and flow steps are listed; shield position is not. |
| Why is that step missing? | The checklist was written before the shield was added to the tool configuration two years ago, and was not revised. | Change record for the shield installation has no linked checklist update. |
The chain stops at a condition the team can change: a checklist that was never updated when the configuration changed. It does not stop at "technician did not reposition the shield." That answer is true but not actionable, because the technician followed the checklist that existed.
Notice what the chain does not do. It does not jump from step 1 to step 5. Each step was confirmed before the next question was asked. If the shield had been found at setpoint in step 2, the chain would have branched to another Machine candidate instead of continuing.
Verify, don't just explain
Explanation says why the cause could produce the symptom. Verification shows that it does. Three levels of evidence, in increasing strength:
| Level | What was done | What it shows | What it does not show |
|---|---|---|---|
| Correlation | Historical data: all low-height lots on tool B were processed after the maintenance date. | The condition and the symptom occur together. | That the condition caused the symptom. Another change at the same time could be responsible. |
| Reproduction | In an approved test setting, the shield is set to its measured displaced position and a test wafer is plated; then set to setpoint and another is plated. Edge height is low in the first case and in specification in the second. | The condition is sufficient to produce the symptom, and removing it removes the symptom. | That it was the only contributor in the original lot. |
| Elimination | Shield restored to setpoint; checklist revised. Tool B edge heights tracked for the following lots stay within specification and match tools A and C. | The fix holds under production conditions over time. | Nothing further, if the tracking period and sample are adequate. |
Correlation alone is where most analyses stop. It is useful as a starting point and weak as a conclusion. In this case, the maintenance event is correlated with the shortfall, but maintenance involved several actions. Correlation does not identify which one.
Reproduction is the step that turns a candidate into a cause. It requires a controlled condition, so it has to be planned: who authorizes it, on what material, and how the result will be judged. This guide does not suggest deliberately running production material under a known-bad condition. Use test wafers, a qualification slot, or an engineering run with approval.
Elimination closes the loop. "We fixed it and it has not come back" is only evidence if the tracking period is long enough and the sample is large enough to have caught a recurrence. State the period and the number of lots. Five lots over two days is a different claim from sixty lots over two months.
If the gauge was not characterized before the analysis, do it before reproduction. Reproducing an 8 µm difference means nothing if the gauge cannot distinguish 8 µm.
Common mistakes
| Mistake | What it looks like | Better practice |
|---|---|---|
| Correlation presented as cause | "Root cause: post-maintenance condition." | Name the specific condition and show reproduction or elimination evidence. |
| Single-cause assumption | The analysis stops at the first confirmed cause. | Ask whether the confirmed cause explains the full extent of the symptom. If 3 wafers failed but the shield affects all 25, why did 22 pass? |
| Stopping at the person | "Technician error." | Ask why the procedure allowed the error. A cause that depends on one person being more careful will recur. |
| Skipping measurement-system confirmation | Gauge assumed adequate. | Confirm repeatability and reproducibility relative to the differences being analyzed. |
| Fix-and-see as the only verification | Change made, no recurrence in a week, case closed. | State the tracking period and sample. Prefer reproduction evidence where it can be obtained safely. |
| Contributing factors recorded as root cause | Bath temperature was 1 °C high, so it is listed as a cause. | Separate factors that worsen the effect from the condition that produces it. Record both; label them correctly. |
| Rewriting the problem statement to match the cause | The statement changes from "edge height low" to "shield displaced" partway through. | The problem statement describes the observation and stays fixed. |
Before you close
- The problem statement describes what was observed and includes location, time, and extent.
- Stratification results are recorded, including the candidates they ruled out.
- The fishbone shows all candidates considered, with pruned branches visible.
- Each 5 Why step has its own supporting evidence.
- The root cause has reproduction or elimination evidence, not only correlation, and the tracking period is stated.
- The measurement system was confirmed before differences were interpreted.
- Contributing factors are recorded separately from the root cause.
Once the cause is verified, the result feeds two documents. The 8D report carries it into permanent corrective action and prevention. The PFMEA for the plating step should be updated: a shield-position cause that was not on the worksheet before now has evidence of occurrence, and the maintenance checklist is a prevention control that was not described.
Verification also tends to leave a quantitative question behind. The reproduction test showed that a 6 mm displacement produces an 8 µm shortfall; it did not show how much displacement the process can tolerate, or whether the sensitivity changes with current density or flow setting. Answering that is not a root cause task. It is an experiment, and it needs factors, levels, and a response defined before any wafer is run. The experimental design guide covers how to set that up. A root cause tells you what went wrong; a DOE tells you how much margin you have against it.
A root cause is a claim about mechanism. Treat it like any other engineering claim: state it, test it, and record what the test showed.
Sources and further reading
- ASQ — Root Cause Analysis. Overview of the purpose and general approach to root cause analysis. (asq.org)
- ASQ — Fishbone (Ishikawa) Diagram. Description of the diagram, its categories, and how to construct one. (asq.org)
- ASQ — Five Whys and Five Hows. Description of the technique and its use within root cause analysis. (asq.org)
- NIST/SEMATECH Engineering Statistics Handbook — Exploratory Data Analysis. Graphical methods for examining data structure before applying statistical tests. (itl.nist.gov)