Insight / Blog
SEM Particle Analysis Automation : Five Days to 10 Minutes
In battery and semiconductor manufacturing, particles too small to see can affect product safety and performance. Irregular cathode-material particles can reduce efficiency, while a single oversized abrasive particle can damage a wafer surface.
Large-area scanning electron microscope (SEM) images must therefore be used to examine the size and shape of thousands of particles. Their physical condition can also provide evidence about process variables such as temperature and pressure.
Background : Limits of Manual Sampling Across Thousands of Particles
Company A, a precision metrology equipment provider, struggled to analyze thousands of particles captured in a single large-area microscopy image. Manual review was slow, and classification standards varied by analyst.
The team often had to rely on sample-based analysis. This created blind spots around overall particle distributions and rare patterns that could indicate process anomalies.
Approach : Clustering, Multimodal Analysis, and AutoML Classification
SURROMIND designed a pipeline that extracts particle shape and size from SEM images and automatically groups particles with similar patterns.
Consensus clustering: compared particle shapes and measurements across multiple criteria to create more stable groups and reduce classification error.
Multimodal clustering: analyzed visual form and measured values together to identify patterns that neither source could capture alone.
AutoML: shortened the process of configuring and evaluating candidate classification models for new datasets.
Evaluation and visualization: presented clustering and classification results in a form that analysts could review and validate.
Results : Full Analysis Reduced from Five Days to About 10 Minutes
| Metric | Manual analysis | SURRO Inspection |
|---|---|---|
| Analysis time | 5 days for a large-area image | Approximately 10 minutes |
| Coverage | Sample-based analysis of selected particles | Full-population analysis across all particles |
| Decision process | Manual selection and classification | Visualized clustering and classification results for expert review |
Operational Impact : Rare-Pattern Detection for Process-Cause Analysis
The objective was not to remove expert analysts. It was to move repetitive particle sorting to AI so experts could spend more time investigating why abnormal patterns appeared.
Previously, analysts spent days comparing thousands of particles by eye. The new workflow analyzed an entire large-area image in about 10 minutes and visualized similar particles as a distribution map.
Quality teams could use the map to understand the overall distribution, isolate unusual or rare particle groups, and focus root-cause analysis on the patterns that mattered most.
