Reducing Casting Scrap Rate: Where to Start When Everything Seems to Be a Factor

The First Question: Is Your Scrap Rate Measurement Accurate?

Before starting process improvement work, verify that the scrap rate you are trying to improve is actually measured correctly. In many foundries, scrap tracking systems mix several phenomena that have different causes and different corrective actions: internal rejects caught at inspection, parts scrapped at the trim press due to handling damage, parts rejected at the customer for incoming inspection failures that escaped internal inspection, and rework parts that are counted as scrap when the rework cost exceeds the part value.

These categories should be tracked separately. A 4% internal scrap rate and 0.5% customer return rate describe a different quality situation than a 2% internal scrap rate and 2% customer return rate, even though the total reject rate is similar. The first situation indicates aggressive internal inspection catching most defects before shipment. The second indicates internal inspection missing roughly half of defective production. The corrective action is different: in the first case, focus on process improvement to reduce defect generation; in the second case, focus on inspection effectiveness improvement alongside process improvement.

For defect-type breakdown within internal scrap, most foundry QC systems have defect code fields that are inconsistently used. "Porosity" covers both gas and shrinkage porosity. "Dimensional" covers both die wear-driven drift and process variation. "Surface" covers cracks, cold shuts, marks, and handling damage. Before analyzing defect cause, verify that your defect codes are being applied consistently and that they provide enough resolution to point to specific process variables.

The Pareto Analysis That Actually Works

Pareto analysis of defect types by frequency is the standard starting point for scrap reduction. The 80/20 rule reliably applies: a small number of defect types account for the majority of scrap volume. Start with the top two or three defect types by scrap piece count and focus all initial effort there.

The typical Pareto in HPDC shows either porosity or cold shut as the dominant defect type, with surface cracks and dimensional issues making up most of the remainder. The proportions vary by alloy, part design, and how tightly dimensional tolerances are set. If your Pareto shows five or six defect types each contributing 15-20%, your defect classification scheme is too coarse - you are aggregating different defect mechanisms under broad category labels.

After identifying the top defect type by frequency, the next Pareto is time-based: plot defect rate for that defect type by shift, by day of week, by die shot count (for die wear correlation), and by operator. Patterns in these plots indicate whether the dominant defect is random (no time pattern), systematic (constant), or periodic (die life cycle, shift change, raw material lot change). Each pattern type points to a different root cause category.

A cold shut rate that is higher on the night shift than the day shift, controlling for all other variables, is pointing to a human process variable - operator-controlled parameters that differ systematically between shifts. Metal temperature management during shift changeover, lubricant application technique, or die cycle start procedures are candidates. This is a different corrective action than a cold shut rate that increases monotonically with die shot count, which points to die wear or progressive die temperature stabilization.

Process Capability vs. Process Centering

Two distinct sources of dimensional scrap require different corrective actions: process capability problems (the distribution is too wide to fit within the tolerance window) and process centering problems (the distribution is centered away from nominal, so parts cluster near one tolerance limit).

Cpk combines both effects: a well-centered process with a wide distribution and an offset-centered process with adequate width both produce low Cpk. But the corrective actions are different. A centering problem is often fixed by adjusting a single set point (die temperature, shot parameters, billet length). A capability problem requires either reducing the variance source (tighter raw material control, better die cooling uniformity) or widening the tolerance (customer negotiation) or both.

Separating Cp (capability without centering adjustment) from Cpk (with centering) reveals which problem you have. If Cp is acceptable (above 1.33) but Cpk is low (below 1.0), the process is inherently capable but running off-center. Find and fix the centering offset. If both Cp and Cpk are low, the variance is too large and cannot be fixed by set point adjustment alone.

In casting operations, common sources of process decentering are: die thermal state that is not yet stable (first few hundred parts after die change or restart run warm or cold), raw material lot changes that shift pour temperature or alloy composition, and lubricant formulation changes that alter fill behavior. These are step-changes that show up as sudden shifts in the control chart rather than gradual drift.

The Data You Probably Don't Have - and Why You Need It

The majority of foundry scrap reduction efforts are constrained by data availability, not by analytical capability. Engineers understand that billet temperature drives cold shut occurrence. They do not have shot-level billet temperature data to analyze. They understand that die temperature distribution affects shrinkage porosity location. They do not have thermal map data for each cycle to correlate with rejection outcomes.

The priority investment for data-limited scrap reduction programs is measurement - adding the sensors and data collection for the process variables that engineering judgment identifies as most likely causal, before starting the statistical correlation analysis. Running correlation analysis on the variables you currently measure is faster but less likely to identify the actual root cause if the causal variable is not in the dataset.

The practical sequence for a scrap rate reduction program starting from limited data: Identify the top defect type by Pareto. Identify the top candidate causal variables by engineering judgment. Confirm which of those variables are already measured and available for correlation. Prioritize adding measurement for variables that are not currently measured and that engineering judgment rates as highly probable causes. Run correlation after at least 4 weeks of new data collection. The correlation analysis then has the variables needed to generate actionable findings.

When to Use DOE vs. Correlation Analysis

Statistical correlation analysis identifies relationships between process variables and defect outcomes from existing production data. Design of experiments (DOE) deliberately varies process variables to measure their effects on outcomes. Each has appropriate applications in scrap reduction work.

Correlation analysis from production data is appropriate when: you have sufficient production data with variable ranges that include the range of interest, the variables of interest vary independently in production (are not correlated with each other), and you want to scan a large number of candidate variables efficiently to identify which warrant deeper investigation.

DOE is appropriate when: correlation analysis has identified a small number of candidate variables but their relative importance is unclear, the variables are correlated in production data (making regression unstable), or you need to characterize interaction effects between variables that production data does not capture.

Many scrap reduction programs skip correlation analysis and go directly to DOE. This wastes experimental resources on variable combinations that could have been eliminated as non-causal by correlation analysis first. The right sequence is correlation analysis on production data to generate a short list of candidate variables, followed by a targeted DOE to quantify the effects and interactions of the candidates.

The ForgePuls process correlation engine automates the first stage of this work by continuously running multivariate correlation analysis between inspection-identified defect outcomes and machine-sourced process variables. The output is a ranked list of variable-defect correlations that directs engineering investigation to the most statistically supported candidates. This does not replace engineering judgment - it focuses it.

For the relationship between inspection data quality and the ability to do this analysis meaningfully, see our article on why generic computer vision models underperform for foundry inspection. The quality of the defect signal from the inspection system directly determines the quality of the correlation analysis it enables.