In semiconductor manufacturing, automated processes make decisions using data configured and maintained by engineers. If the rules behind that data are unclear, inconsistent, or incorrectly applied, the resulting behavior can be difficult to anticipate. Business rule discovery offers a way to examine process data and make the rules reflected in day-to-day operations more visible.
For automation engineers and process-control teams, the goal is not simply to build another predictive model. It is to identify patterns that can be expressed as understandable rules, then use those rules to focus expert review on suspicious configurations and exceptions. A data-mining approach tested on semiconductor measurement-step data illustrates how this can support process understanding and practical model improvement.
Why business rule discovery matters in automated manufacturing
Business processes are often described through their control flow: which activities occur, and in what order. But process behavior also depends on data. Data can determine which path a process takes, what action an automated system performs, and how a particular production condition is handled.
In a knowledge-intensive environment, those decisions may encode operational rules that are not fully documented. The expertise may reside with engineers, or be embedded in the configuration of a manufacturing execution system (MES). When a process changes, or when configuration is transferred between teams, gaps in that knowledge can contribute to errors or inconsistent decisions.
Discovering the rules already reflected in process data can help organizations:
- Build a clearer picture of how an existing process behaves.
- Identify instances that do not match common patterns for expert investigation.
- Support knowledge capture by making operational rules easier to review.
- Provide a basis for checking and improving process models.
From event logs to data-driven decision rules
Process mining commonly uses event logs to examine how activities unfold across process instances. It can support process discovery, conformance checking, and model enhancement. Business rule discovery adds a complementary focus: rather than looking only at activity order, it examines how data attributes relate to decisions made within a process.
A decision tree is useful for this purpose because its structure can be read as a series of tests. Each path from the root of the tree to a leaf can be translated into an “if–then” rule. The conditions form the “if” side, while the leaf’s classification forms the “then” outcome. This creates a model that domain experts can inspect and discuss.
The method described uses the C4.5 decision-tree algorithm, which selects attributes for splits using gain ratio. Pruning is applied to limit unnecessary complexity: tree growth can stop when too few instances remain at a node, and reduced-error pruning can remove branches when doing so does not affect prediction accuracy. These steps are intended to make the resulting rules more manageable for review.
Using rule coverage and accuracy to flag exceptions
Each extracted rule can be assessed using two measures. Coverage indicates the proportion of all records that meet the rule’s conditions. Accuracy indicates the proportion of records meeting those conditions that also match the rule’s outcome.
These measures help distinguish different kinds of findings. A rule with low coverage may describe a rare case or an unusual configuration, so the records it covers can merit closer inspection. A high-coverage rule with strong accuracy expresses a recurring pattern. Records that meet its conditions but produce a different outcome can be flagged as potential exceptions.
A rule with weak accuracy may indicate that the available attributes do not adequately explain the decision. Rather than treating every mismatch as an error, engineers can use the result as a prompt to investigate the process context and consider whether a relevant variable is missing. Expert review remains essential: a statistical pattern identifies where to look, but does not by itself establish why a deviation occurred.
Applying the approach to semiconductor measurement steps
The industrial application focused on measurement steps in a semiconductor manufacturing route. A product may pass through more than 200 production steps, and measurement steps can affect both quality and productivity. These steps are characterized by measured parameters and Statistical Process Control (SPC) chart types, as well as actions associated with conditions such as Out Of Specifications (OOS) or Out Of Control (OOC).
Because the measurement steps were managed through an MES, the configuration informed fully automated process execution. The analysis used 122,984 instances from a High Mix Low Volume production line. The selected data included step, parameter, equipment information, SPC-related attributes, and OOS and OOC actions. The resulting decision-tree rules were reviewed with SPC engineers rather than judged primarily by predictive performance.
In one example, a rule had 8,400 records matching its conditions, with 8,397 matching the expected outcome. The three exceptions were sent to experts for analysis, which led to the deletion of three control charts. In another correction, a HOLD LOT decision was changed for several charts. The study reports that the work resulted in 14 engineering change notices, each corresponding to a correction to one or more SPC chart definitions.
Turning discovered rules into a review workflow
For plant teams considering a similar approach, the practical value depends on connecting analysis to engineering review. A workable sequence is to select process attributes with domain experts, train and simplify a decision tree, translate its paths into rules, and examine coverage, accuracy, and exceptions. Engineers can then validate findings against the process context before making configuration changes.
That distinction matters in automated production. An unusual record may be a genuine model error, a legitimate special case, or evidence that the data does not capture an important condition. Treating every deviation as a fault risks incorrect changes; treating none as actionable leaves potential issues undiscovered. Rule-based analysis helps prioritize attention while keeping decisions with the relevant experts.
Conclusion: make process behavior easier to inspect
Business rule discovery can help semiconductor manufacturing teams connect process data with the decisions encoded in automated workflows. Interpretable decision trees turn recurring patterns into reviewable rules, while coverage and accuracy help identify records that warrant investigation. Used with expert validation, the approach can strengthen process understanding, support knowledge capitalization, and guide targeted improvements to measurement-step models and SPC actions.