Intelligent and automatic characterization of ignitable liquid residues by using total ion spectrum and machine learning

The identification and classification of ignitable liquid residues (ILRs) are among the most challenging steps in a fire investigation. Traditionally, fire debris samples are analyzed by GC–MS according to ASTM E1618 guidelines, where ILRs are determined via chromatographic pattern comparison with reference samples and classified according to ASTM classes. This complex and partially subjective task relies heavily on an experienced analyst and does not lend itself to automation. This study aims to provide alternative methods for the automatic data classification based on the total ion spectrum (TIS) combined with machine learning (ML) algorithms. We evaluated and compared thermal desorption of ILRs from activated charcoal strips (ACS samples) and direct headspace analysis of different fire debris samples to develop automated discrimination of different ILRs. Hierarchical cluster analysis (HCA) and supervised ML algorithms like linear discriminant analysis (LDA), support vector machines (SVM), and random forest (RF) were employed. The samples tended to initially cluster according to the presence/absence of ILRs followed by the type of ILR. Three datasets were tested (1) ILRs presence/absence, (2) ILR classification, and (3) combination of both. For ACS samples, all the datasets achieved 100% classification accuracies. For direct headspace analysis, dataset (1) resulted in 100% accuracies when the three algorithms were used. Models using dataset (2) and (3) achieved less than 100%. A prototype web-based application for the automatized discrimination of ACS samples was created using the SVM model. The application of ML algorithms demonstrated the effectiveness of the developed methods as supplementary to traditional ASTM method. These methods provide significant potential for automating the classification of ILRs therefore they can be useful to the experts in interpreting fire debris data.