Convolutional neural networks (CNNs) are inspired by the visual cortex of the brain. In this work, CNNs, are applied to classify ground truth samples as positive or negative for ignitable liquid residue (ILR+ and ILR−, respectively).
Known ground truth samples included laboratory-generated fire debris samples, neat ignitable liquids (ILs), single-substrate (SUB) burned samples and computationally generated (in silico) training samples. The images were generated from the total ion spectra for both training and test datasets by applying a wavelet transformation. The training set consisted of 50,000 in silico-generated fire debris samples. The probabilities generated from the CNN are used to calculate the likelihood ratios. These likelihood ratios were calibrated using logistic regression and the empirical cross-entropy (ECE) plots were used to investigate the calibration of the probabilities of the presence of ILRs (i.e., probability of belonging to class ILR+). The performance of the model was evaluated by the area under the receiver operating characteristic plots (ROC AUC). The ROC AUC for the laboratory-generated fire debris samples and the combined IL and SUB samples was 0.87 and 0.99, respectively. The CNNs trained on in silico data did significantly better predicting the classification of the pure IL (ILR+) and SUB (ILR−) samples. Nonetheless, the classification performance for laboratory-generated samples was sufficient to aid forensic analysts in the classification of casework samples. (Publisher abstract provided)