Successful Machine Learning Applications in Forensic Sciences

Omer Kaspi, Chemistry, Bar Ilan University, Rosh Haain, Israel (omerkaspi@gmail.com)

In recent years, models derived with the help of machine learning (ML) technologies have been widely and successfully used in various research fields including chemistry, biology, toxicology, and more recently material sciences and forensic sciences. Briefly, ML uses a dataset of labeled samples each characterize by a set of features (i.e., descriptors) to train a model. Depending on context and on the question at hand, these models can identify abnormalities in patterns, classify a sample into one of pre-determined groups or predict numerical values for a variety of end-points.

In this work we present the application of ML to two problems of forensic relevance. Both projects were performed in collaboration with the Israeli Division of Identification and Forensic Sciences (DIFS). In the first project we have developed models that can reliably (with ~85% accuracy) classify glass fragments taken from cars’ windshields into one of ten different car manufacturers, based on the experimentally measured elemental composition of the glass. Furthermore, we have demonstrated the benefits on model performances obtained by combining measurements from different labs and from different analytical techniques. This may pave the way to international collaboration between law enforcement agencies. In the second project, we have developed a model for the classification of ignitable liquids into petroleum distillates, gasoline and a group of materials that does not belong to either category, with above 95% accuracy. As part of this work we have developed a new method for generating synthetic data thereby overcoming barriers to model development resulting from small datasets. We believe that these, and similar models that could be developed for additional forensic domains, could pave the way towards better exploration of forensic evidence en route to solving crimes.