Machine learning in particle analysis: new standards for technical cleanliness
Why traditional methods are reaching their limits
The precise segmentation and classification of particles is essential for quality assurance in production. However, conventional image processing methods reach their limits with transparent, overlapping or low-contrast particles. Mica particles in particular, which are used in e-mobility as a fire protection material, are often incorrectly classified as shiny metallic.
Machine learning as a key technology
By using machine learning algorithms, in particular the random forest classifier, particles can be classified much more reliably. The advantages:
- Higher classification accuracy
- Reduced manual post-processing effort
- Robust results even with complex particle structures
CleanControlling shows: Mica classification significantly improved with ML
In a real use case, CleanControlling demonstrated how ML reduces the misclassification of mica particles by over 94%. Even smaller particle sizes (<200 µm) were reliably detected - a clear advantage over conventional methods.
How the training of the ML model works
The ML model is taught in several steps:
- Image acquisition and post-processing of the particle filters
- verification using SEM-EDX and IR analysis
- selection of representative image sections
- annotation (labeling) of the particle types
- training and validation of the mode
- iterative fine-tuning until optimal prediction accuracy is achieved
The trained model is integrated into the ZEISS ZEN core TCA software and can be further trained by users.
Conclusion: machine learning increases efficiency and quality
The use of machine learning in particle analysis offers enormous advantages:
- Significant reduction in misclassifications
- More efficient processes in particle typing
- Higher reproducibility and objectivity
For companies that rely on the highest standards of technical cleanliness, the use of ML is a decisive competitive advantage.
Newsletter registration