Evaluation of Microscopic Foreign Matter in CMP Slurry Using Dynamic Image Analysis and Machine Learning

Significance of the topic

Chemical Mechanical Planarization (CMP) slurries for semiconductor polishing must be free of microscopic foreign matter because such contaminants can produce defects on wafer surfaces. Detecting and characterizing trace micron-scale and coarse particles within high-concentration colloidal silica slurries is therefore critical for process control, filter design, and yield improvement. Dynamic image analysis (DIA) combined with machine-learning classification provides a route to identify and quantify particle morphologies at near-undiluted concentrations where traditional light-scattering methods struggle to detect rare large particles.

Objectives and study overview

This application study evaluated the capability of the iSpect DIA-10 dynamic particle image analysis system to detect and characterize trace micron-scale foreign matter in a high-solid-content (≈30 wt%) colloidal silica CMP slurry. Key aims were to (1) measure micron and larger particles without heavy dilution, (2) obtain morphological information from particle images, and (3) apply machine-learning clustering to classify particle types for improved source identification and mitigation strategies.

Instrumentation Used

• iSpect DIA-10 dynamic particle image analysis system employing a microcell flow channel and telecentric optics (high imaging efficiency ≥ 90%).
• SALD-7500nano laser diffraction nanoparticle size distribution analyzer (used for comparison; measurements required 1000× dilution).

Methodology

Undiluted polishing-grade colloidal silica (approx. 30 wt% solids, water dispersant) was analyzed by DIA to capture images of particles flowing through a microcell. For reference, laser diffraction (SALD-7500nano) measured the sample after 1000-fold dilution and indicated a monodisperse distribution with median diameter ≈88 nm. DIA settings during the study included a frame rate of 8 fps, imaging efficiency ~96.5%, luminance 249, binary-image threshold 90%, pump volume 250 μL, and background correction enabled. Individual particle images were extracted to compute area-based diameter and derived sphere-equivalent volumes. Machine-learning classification used UMAP for dimensionality reduction and HDBSCAN for clustering to group particles larger than 10 μm into distinct morphological populations.

Main results and discussion

Key quantitative findings:

• DIA on the undiluted sample detected 3,315 particles per mL in the micron size range; the largest detected particle had an area-equivalent diameter of ~73 μm.
• The summed sphere-equivalent volume of all detected particles corresponded to a volume concentration of 2.337 ppm, much lower than concentrations typically used for laser diffraction/scattering measurements (hundreds of ppm), explaining why diffraction methods would miss these rare coarse particles.
• At a 1000-fold dilution, both DIA and laser diffraction detected only very few coarse particles, demonstrating that near-undiluted measurement is preferred for reliable detection of trace foreign matter.

Morphological classification and interpretation:

• Particles >10 μm (which accounted for >90% of total particle volume) were classified into six clusters via UMAP + HDBSCAN. Clustering used morphological descriptors such as area-based diameter (ABD), aspect ratio (AR), circularity, and average brightness.
• Clusters labeled a, b, and d showed low aspect ratios (elongated shapes). Clusters a and d also exhibited higher average brightness, suggesting thin, flake-like particulates with low thickness in the optical observation direction.
• Other clusters represented more compact or intermediate morphologies; differences in brightness also provide clues on refractive-index contrast, thickness, or surface roughness.

Practical implications from morphology:

• Knowing that several populations are elongated indicates the importance of selecting filter pore geometries based on the minor axis (not just nominal pore diameter) to ensure removal.
• High-brightness flake-like particles imply different origins (e.g., polymer flakes, abrasives fragments, or debris) compared with compact agglomerates; this informs targeted mitigation (filtration, upstream contamination control, or slurry formulation adjustments).

Benefits and practical applications

The DIA + microcell approach demonstrates several practical advantages for CMP slurry quality control:

• Direct imaging at near-undiluted or low-dilution conditions enables detection of rare but critical coarse particles that would be undetectable by standard scattering methods.
• Telecentric optics and high imaging efficiency reduce missed-particle events, improving confidence in low-count detections.
• Image-derived morphological metrics and machine-learning clustering support source identification, filter selection, and process optimization by distinguishing flake-like, elongated, and compact particle types.

Future trends and potential applications

Adoption of DIA combined with advanced analytics is likely to grow in CMP and other high-purity process industries. Emerging and potential trends include:

• Integration of real-time image analysis and automated clustering to provide inline contamination alerts and trend monitoring in manufacturing.
• Enhanced classification models (deep learning) trained on comprehensive libraries of particle images to improve root-cause identification and reduce manual review.
• Correlation of particle morphology with performance outcomes (e.g., defect maps on wafers) to prioritize control measures that most reduce yield loss.
• Optimization of filtration and contamination control hardware based on quantitative minor-axis and shape distributions instead of only nominal particle size metrics.

Conclusion

Dynamic image analysis using the iSpect DIA-10 microcell permits reliable detection and morphological characterization of trace micron-scale and coarse foreign matter in high-concentration colloidal silica CMP slurries without the need for extensive dilution. The method detected particle populations and sizes that would be missed by conventional laser diffraction after the dilution required for those techniques. Coupling image-based metrics with machine-learning clustering yields actionable classification of particle types, aiding filter selection and contamination mitigation strategies. Overall, DIA is an effective tool for CMP slurry QC where trace coarse particles critically impact wafer yield.

References

1. Syuhei Kurokawa: The Overview and Future Prospects for Planarization CMP Technology, Journal of the Japan Society for Precision Engineering, Vol. 84, No. 3, pp. 213–216 (2018).