Toward Developing Techniques─Agnostic Machine Learning Classification Models for Forensically Relevant Glass...

Area of Science:

• Forensic Science
• Materials Science
• Analytical Chemistry

Background:

• Glass fragments are crucial forensic evidence for origin determination.
• Machine learning (ML) models require large, diverse glass fragment databases for reliable classification.
• Standardized analytical techniques are ideal but often unfeasible across different laboratories.

Purpose of the Study:

• To investigate the feasibility and methods for combining glass fragment measurement data from different laboratories and techniques for ML-based classification.
• To establish rules for successful data combination and assess their impact on model performance.
• To improve the performance of existing techniques like SEM-EDS through data integration.

Main Methods:

• Evaluation of various elemental analysis techniques (PIXE, LA-ICP-MS, SEM-EDS, PIGE, INAA, PGAA) for generating lab-specific ML classification models.
• Determination of data combination rules across different laboratories and analytical methods.
• Comparative analysis of model performance with combined versus lab-specific data.

Main Results:

• Each elemental analysis technique can produce effective lab-specific ML classification models.
• Combining PIXE and LA-ICP-MS data improved model performance by approximately 10-15%.
• Combining PGAA data with other techniques yielded poorer results compared to lab-specific models.
• Replacing SEM-EDS measurements for Fe and Ca with PIXE measurements significantly enhanced SEM-EDS model performance.

Conclusions:

• Data from different laboratories and elemental analysis techniques can be successfully combined to create improved, lab-agnostic ML models for glass fragment classification.
• Established rules for data combination are critical for model performance; poor combinations lead to decreased accuracy.
• This approach facilitates the creation of unified databases and fosters international collaboration among forensic laboratories, enhancing the utilization of glass fragment evidence.