Fdc (fault Detection and Classification)

Fault detection and classification (FDC) turns equipment signals into evidence that a semiconductor process run behaved abnormally, then helps distinguish possible fault types. Pressure, temperature, gas flow, RF power, valve position, and other signals describe what the tool did while a wafer was being processed. FDC compares that behavior with qualified expectations so engineers can investigate sooner and contain potentially affected material.

Detection is not diagnosis. A pressure deviation might reflect a gas-delivery problem, a pumping problem, a changed recipe, or a faulty sensor. Detecting the deviation establishes that something needs attention; classification proposes a category supported by the available evidence. Neither step automatically proves the physical root cause or establishes that the wafer is defective. Inspection, metrology, equipment checks, and process history can provide the missing evidence.

Compare the right part of the recipe. Normal operation can include pump-down, gas stabilization, plasma ignition, processing, and purge, with very different expected signals in each phase. A single limit applied to the entire run can mistake a normal transition for a fault or hide a process-stage deviation. Step identifiers and event timing allow traces to be aligned and compared within the correct phase. Transition duration can itself be a useful monitored feature.

Useful features preserve the mechanism of interest. A step mean can reveal a sustained offset, but it can dilute a short spike. Maximum value, slope, settling time, oscillation, or a residual relative to an expected trace may be more informative for particular failure modes. Sampling and filtering must preserve the events being sought. A fast data stream is not automatically useful if timestamps are misaligned, units change, or a sensor is stuck. Missing samples and invalid sensor states should be handled explicitly rather than interpreted as healthy process data.

Some abnormalities live between sensors. Two readings can each be inside their individual limits while their combination is unusual. For example, signals that normally rise together may suddenly move in opposite directions. Multivariate monitoring evaluates such relationships. A Hotelling T-squared statistic summarizes distance in a covariance-aware space; principal component analysis (PCA) represents patterns of correlated variation. These methods can flag a departure that separate limits miss, but a large score alone does not identify a failed component.

The reference must match the operating context. Product, recipe revision, chamber, and maintenance state can change the expected signature. A model built on one qualified population should not silently treat every other population as equivalent. Likewise, a reference should represent acceptable operation rather than simply the most recent runs: continually learning from degraded behavior can normalize the very fault the system should detect. Changes to baselines and models need versioning and validation.

Classification requires evidence and an unknown category. Rules can associate known signal patterns with candidate fault families. Statistical or machine-learning classifiers can use labeled historical examples, but their confidence depends on the relevance and quality of those labels. New faults may not resemble the training examples. An explicit unknown or uncertain outcome is preferable to a confident but unsupported component diagnosis. Signal contributions help direct investigation, not prove causation.

The response depends on timing and severity. Some checks run during processing; others use completed traces after the run. Approved responses may include an alert, a material hold, disabling further starts, or a qualified equipment action. FDC is not a replacement for equipment safety interlocks. Wafer genealogy identifies material processed during the suspect interval, including lots already downstream. Detection delay determines how much material may need review before a cause is confirmed.

FDC and APC have complementary jobs. APC compensates for qualified process variation through controlled recipe changes. FDC can prevent abnormal runs from being treated as ordinary drift when updating that controller. Otherwise, a controller may compensate for a failing tool or learn a biased model. A clean FDC result still does not guarantee wafer quality: the fault may be invisible to the installed sensors, so physical metrology and inspection remain necessary.

Performance is a trade-off, not a perfect alarm count. Tight thresholds can detect smaller changes but increase nuisance alarms; loose thresholds can miss harmful events. Evaluate missed faults, false alarms, detection delay, and practical response burden on representative data. Preserve time order when testing models, and reassess performance after process changes. The useful outcome is earlier, trustworthy intervention with a traceable reason for the action.

Sources: NIST Engineering Statistics Handbook, Principal Components Control Charts and What Are Control Charts; Applied Materials, E3 Equipment and Process Control; Onto Innovation, Smart Manufacturing Software.