How a high-volume advanced packaging manufacturer achieved >99% defect classification accuracy and reduced review time from ~30 minutes to ~5 minutes per lot with SixSense AI.

A Taiwan-based advanced packaging manufacturer was rapidly ramping production of high-density Redistribution Layer (RDL) devices.

The inspection environment was already operating at significant scale:

• 20M+ inspection images per year
• Multiple AOI machines, with additional tools planned
• 10+ defect classes
• High-volume production continuing to ramp

RDL inspection is particularly challenging because semi-transparent materials and dense, repetitive patterns can make subtle defects difficult to distinguish from normal structures.

Challenges

1. Manual review could not scale with production

Before SixSense, operators used to manually review and classify defect images.

As production ramped, the customer faced:

• Increasing review volumes
• Limited availability of operators
• Higher cycle time as inspection volume increased
• The need to add more operators as production expanded
• More false alarms as inspection criteria were tightened to protect product quality

The customer needed inspection throughput to scale without review headcount scaling with it.

2. Misclassified Defects Lead to Wrong Root Causes and Actions

Scale was only part of the problem. Across a sample of approximately 100,000 images, 25–30% showed inconsistent labelling between operators.

Images originally classified as Contamination were subsequently identified by engineers as different defect classes, including Interconnect Bridging, Surface Irregularity and Shape Deviation.

Without accurate, consistent defect classification, manufacturers cannot reliably automate root cause analysis or corrective actions. Reliable classification is the foundation for autonomous manufacturing.

3. Subtle defects created risk of escape

RDL's semi-transparent material made the visual problem itself difficult. Defect appearance could vary with the underlying structure and imaging conditions.

One real example was a tiny discontinuity in an RDL routing pattern. The surrounding structure looked normal and repetitive. Only a small section of the routing was broken, and the discontinuity blended closely into the surrounding pattern.

Such hard to spot subtle defects lead to critical failures downstream and can be high risk escapes if missed during high-volume manual review.

Example from the project: a subtle break in an RDL routing pattern that closely resembles the surrounding good structure.

4. Rare defects made conventional model development difficult

The customer also had long-tail defect classes where only a small number of examples existed. For example, UBM residue and Big Bump defects were rare, giving the team limited examples from which to teach the model.

This created a familiar manufacturing-AI problem: Millions of images existed, but the images that mattered most for training were often the hardest to find.

Solution

SixSense addressed these challenges through an integrated inspection workflow that combined automated classification, data quality improvement, explainable model diagnosis and data-efficient training.

1. Automate Classification, Review Only What Needs Attention

To address the growing manual review workload, SixSense introduced AI-based defect classification across the customer's AOI inspection images.

Instead of requiring operators to manually classify every image, the system automatically classified defects and enabled operators to focus on images requiring further review.

This allowed the customer to:

• Automate repetitive classification decisions
• Reduce the volume of images requiring manual review
• Accelerate lot-level review and disposition
• Support increasing production volumes without proportional growth in review headcount

The result: 98.7% of inspection decisions automated, reducing lot review time from ~30 minutes to ~5 minutes.

2. Clean inconsistent human labels

To address inconsistent operator decisions, SixSense used AI-assisted review and model analysis to identify images where the existing labels deserved another look.
Engineers could investigate these cases rather than manually rechecking the entire historical dataset.

This helped establish more consistent defect definitions and improve the quality of the examples used for training.

The result: Data quality improvement alone increased model accuracy from 76.5% to 95.7%.

The key finding was that the bottleneck was not simply the quantity of labelled data, but whether the examples were representative and consistently labelled.

3. Diagnose Why Subtle Defects Are Misclassified

To improve classification of visually subtle defects, SixSense provided engineers with tools to understand why a model made an incorrect prediction, rather than relying on blind retraining.

Two capabilities were particularly valuable:

When the model struggled with a particular image, SixSense gave engineers tools to understand why.

Similar Images helped engineers find visually related training examples to identify missing defect variations or inconsistent labels.

Heatmaps revealed which image regions influenced the model's predictions, helping engineers understand whether the model was focusing on the actual defect or irrelevant surrounding structures.

For example, when investigating a subtle RDL discontinuity, engineers could determine whether:

• The model focused on the discontinuity or surrounding circuit patterns
• Similar defects with different appearances were adequately represented in training
• Incorrect labels or unseen variations contributed to the misclassification

Based on these findings, engineers could correct labels, add targeted examples or improve the dataset's coverage of difficult defect variations.

The result: A systematic, evidence-driven approach to improving classification of subtle defects, rather than repeated trial-and-error retraining.

4. Learn rare defects with limited examples

To address the shortage of examples for rare defects, SixSense combined intelligent data selection with semiconductor foundation models and targeted augmentation.

Unique Images automatically surfaced diverse, representative images from large inspection datasets, helping engineers:

• Remove repetitive examples
• Identify different visual variations within the same defect class
• Discover rare examples that might otherwise be overlooked
• Build representative datasets with significantly less manual selection

For defect classes with very few examples, targeted data top-ups and augmentation helped improve training coverage without waiting for large volumes of new production images.

Combined with SixSense's semiconductor foundation models, this enabled the system to learn useful defect representations from limited customer-specific examples.

The result: SixSense trained the model using only 7.5% of the available customer images.

Results

The SixSense AI solution demonstrated high classification performance while significantly reducing manual review effort.By automating most classification decisions, SixSense reduced the time operators spent reviewing each lot while maintaining high classification accuracy.

The approximately 6× improvement in review throughput provides a path to absorb production growth without proportionally increasing manual review headcount.

Business Impact

The customer started with a difficult combination of rapidly increasing inspection volume, constrained labor, inconsistent human decisions, subtle defects and limited examples of rare defect classes.

SixSense demonstrated a different way to scale inspection:

• Scale throughput without proportional growth in review headcount
• Standardize decisions with consistent, repeatable defect classification
• Protect quality with >99% accuracy and 98.7% automated decisions
• Accelerate lot review with ~6× faster review throughput
• Reduce training effort by using only 7.5% of available customer images
• Build a foundation for autonomy through reliable defect classification that supports downstream root cause analysis and corrective actions

For a production environment generating 20M+ inspection images annually, the impact goes beyond automating one classification task.

SixSense gives manufacturers a way to scale inspection at the speed of production, without scaling manual review at the same rate.