Casting and Molding Defect Data: Sourcing From Foundries and Molders

What makes these processes a separate buying problem

In machining, a defect is usually a local departure from a nominal surface. In casting and molding, defects are products of a process: molten material fills a cavity, cools, shrinks and ejects. The same tool makes thousands of near-identical parts, and its defects drift with wear, temperature and maintenance. An archive from such a line is highly repetitive, and its labels are tied to process events as much as to appearance.

That has two consequences. First, defect classes are process-specific and need a process-aware taxonomy. Second, the strongest signal for splitting and generalization is the tool, not the image.

Surface and internal defects

Surface forms can be collected from ordinary machine-vision archives; see in-line machine vision archives and metal surface defects. Internal forms typically need the specialist modality discussed in X-ray and CT inspection images, and a buyer who wants both should expect two supply paths that must be joined by part identifier.

What public X-ray work offers

GDXray includes a castings group within its 19,407 X-ray images, with annotations as boxes or labels, and is free for research and education (Mery et al., 2015). That makes it a reference for how internal casting indications are annotated.

It covers neither injection-molded parts nor your alloy, geometry or detector, so use it only as a capture and labeling reference. Because its terms are research and education, check them before any commercial use; this is not legal advice.

Process context is part of the label

Ask whether images can be connected to process records, because that join turns a picture into an explanation.

• Cavity or die number. A defect that appears in cavity 4 only points to a tool, not a material problem.
• Shot or cycle ID and timestamp. Allows alignment with process-monitoring data when the supplier holds it.
• Material lot or heat. Separates raw-material variation from tool variation.
• Tool maintenance events. Marks regime changes: a polished or replaced tool resets the defect pattern.
• Operator or shift. Useful for human-graded labels.

A buyer does not need all of these. A buyer who wants to model root cause needs most; one who wants only to classify visible flash needs the cavity and lot. The wider story on linking images to what happened next is in inspection images linked to outcomes.

Splitting and generalization

Images from one cavity on one day are near-duplicates. If those land on both sides of a split, test results measure memorization of a tool. Common holdout units, depending on the claim, are tool or cavity, production week, material lot and plant. A model validated only on parts from the tools it trained on says little about a new tool; domain shift across lines goes further. Labels also need a common vocabulary; see defect taxonomy and label schema.

Handling after molding can add marks that look like process defects, such as scuffs from conveyors or stacking; handling-induced damage covers how to keep that distinction in labels.

Illustrative record layout for a molded-part package

# Illustrative only - synthetic ids and values
package: molded_part_defects_v0
process: injection_molding
part_family: housing_type_Q
records:
  - part_id: p_0091823
    tool_id: tool_12
    cavity: 4
    shot_id: shot_77201
    material_lot: resin_lot_K9
    images:
      - {file: p_0091823_top.png, view: top, lighting: dome}
      - {file: p_0091823_side.png, view: side, lighting: low_angle}
    labels:
      - {defect: sink_mark, box: [x, y, w, h], labeler: grader_C}
    disposition: scrap     # accept | rework | scrap
    internal_check: none   # or xray_ref_id
split_key: tool_id
excluded: robot actions, joint states, forces

A record like this lets a buyer choose to hold out a tool, a lot, or a time window. The identifiers are what carry the value.

Quality and acceptance pointers

• Cavity balance. Count images per cavity and per tool; one dominant cavity would distort training.
• Defect-class plausibility. Check that class names match how the process actually fails, not a generic list.
• Label timing. Confirm whether labels were assigned inline, at end of line or after customer return.
• Normal samples. Ask for good parts from the same tools and periods, since contrast with defects drives learning.
• Format. Confirm lossless images for any defect smaller than a few pixels.

The acceptance values are for the buyer to set; none is universal.

What this data cannot support

Casting and molding defect images describe the state of manufactured parts. They do not contain the motion of an operator removing a part, trimming a gate, placing a core or loading a press, so they cannot train a humanoid to perform those tasks; control learning needs robot or teleoperation data. Nor do they show service-life behavior: a part that passed final inspection may still fail in the field, and that signal sits in warranty and field failure data if it exists at all.

Foundry and molding archives are also skewed toward good parts. Severe defects are caught early, rare modes are sparse, and costly internal imaging is often sampled rather than universal. A claim of coverage across defect types should be verified class by class.

Questions to ask suppliers

1. Which processes, alloys or resins, and part families does the archive cover?
2. Are images linked to cavity, tool, shot and material lot?
3. Which defects were labeled from visible inspection, and which from X-ray or CT?
4. Was a second inspector involved, and how were disagreements resolved?
5. Did the tooling, camera or lighting change during the archive's span?
6. Are good parts from the same tools and periods retained?
7. Do dispositions include rework and customer returns, or only final scrap?
8. What does the customer agreement say about reusing images of customer-owned parts?
9. Are images retained in original resolution or compressed copies?

Permissions and provenance

Molders and foundries commonly produce customer-owned geometries under supply agreements. Images may reveal part shape, logos or material specifications. Ask who owns tool and part designs, whether the contract restricts records, and whether regulated programs are involved. This is not legal advice, and removing part numbers does not on its own settle a rights question. See confidentiality in defect data for the broader checklist.

Discuss your requirements

If you need casting or molding defect data, describe the process, defect classes, imaging modality and permitted uses to SourceX. SourceX reviews the requirement and follows up about potential sources among US businesses; a request does not guarantee a match. Return to the inspection data hub or the humanoid training data guide, or see the manufacturing quality records catalog entry.

Explore sourcing and licensing options with SourceX

Sources

• GDXray: The Database of X-ray Images for Nondestructive Testing — castings group and annotation form as a public reference; licensing scope for research and education.