Artificial intelligence and thermodynamics help solving arson cases

Arson investigations often lack traditional evidence like DNA or fingerprints due to fire damage. One critical piece of remaining evidence is the presence of accelerants, often gasoline. Linking a gasoline sample from a fire scene to a suspect requires comparing the weathered sample from the crime scene with an unweathered sample from the suspect. This is challenging because gasoline's composition significantly changes upon weathering due to evaporation, a complex process involving hundreds of components with varying evaporation rates and potentially influenced by factors such as temperature, water from fire suppression, adsorption, and microbial degradation. Existing methods, like PCA, LDA, CVA, HCA, and likelihood ratio approaches, have limitations: they often handle a limited number of components, struggle with highly weathered samples, require extensive experimental data, and fail to predict the original (unweathered) composition crucial for suspect comparison. This paper introduces a novel method to overcome these limitations.

Previous research on analyzing weathered gasoline samples in arson investigations has relied on various statistical and modeling techniques. Statistical methods like principal component analysis (PCA), linear discriminant analysis (LDA), canonical variate analysis (CVA), hierarchical cluster analysis (HCA), and covariance mapping have been employed to identify patterns and relationships within the data. However, these methods often lack the ability to reconstruct the original composition of the weathered gasoline, making it difficult to directly compare it with unweathered samples obtained from suspects. Other studies have incorporated explicit evaporation models based on Raoult's law or gas chromatographic retention data, but these approaches often suffer from limitations in handling the complexity of gasoline mixtures and highly weathered samples. The proposed method improves upon these existing approaches by combining machine learning with detailed thermodynamic modeling, enabling a more accurate and comprehensive analysis of weathered gasoline samples.

This research proposes a novel method combining machine learning, advanced thermodynamic modeling, and quantum chemical calculations to accurately predict the original (unweathered) composition of a weathered gasoline sample. The approach focuses on the ideal scenario where the composition change is solely due to evaporation, neglecting other factors like adsorption or microbial degradation for model development and testing. The thermodynamic model is based on the gamma-phi approach for vapor-liquid equilibrium (VLE) calculations and requires solving a differential equation (Eq. 1) numerically, which is essentially a modified Raoult's law. This equation accounts for the non-ideality of the gasoline mixture using activity coefficients derived from COSMO-RS, a quantum chemical model. The activity coefficients are crucial because they represent the deviation from ideal behavior caused by intermolecular interactions within the complex gasoline mixture, particularly affecting the evaporation behavior of polar components. Artificial neural networks (ANNs) were trained on a substantial dataset of 459 gasoline samples obtained from the Netherlands Forensic Institute (NFI), with the model focusing on the 60 most abundant components found in arson samples. The ANNs predict the degree of weathering (evaporation percentage) based on the composition of a given sample. The integration of the thermodynamic model with the ANN's weathering prediction allows for the backward tracing of the weathered sample to its original, unweathered composition. The model's accuracy was evaluated by comparing the predicted composition with the actual composition, focusing on the deviation between the predicted and actual values for each component. The model utilizes a deep neural network with five hidden layers, and the hyperparameters were chosen to ensure optimal prediction accuracy.

The study demonstrates the successful application of the proposed method in accurately predicting the unweathered composition of gasoline samples. The thermodynamic model, incorporating activity coefficients from COSMO-RS, accurately represents the evaporation behavior of both nonpolar and polar gasoline mixtures, highlighting the importance of considering non-ideal behavior. The ANN model effectively predicts the degree of weathering, with a deviation of around 3% for samples weathered up to 80%. Back-tracing using the ANN and the thermodynamic model yields accurate predictions of the original composition, with a maximum deviation of approximately 4% for even the most volatile components. This is significantly lower than the natural intervariation between gasoline samples, making forensic comparisons more effective. The results indicate that the inclusion of volatile components, previously problematic, is now possible, substantially improving the effectiveness of forensic gasoline comparisons in arson cases. The analysis of a seven-component nonpolar mixture and a four-component mixture containing ethanol demonstrated the model's ability to account for the non-ideality of complex mixtures and the influence of polar compounds on evaporation. Figure 2 illustrates these effects, comparing ideal and non-ideal evaporation behaviors. Figures 3 and 4 showcase the performance of the ANN in predicting the evaporation percentage and reconstructing the original composition of weathered samples.

This research addresses the significant challenge of comparing weathered and unweathered gasoline samples in arson investigations. The successful integration of thermodynamics, quantum chemistry, and machine learning provides a robust and accurate method for predicting the original composition of weathered gasoline samples. The findings demonstrate the limitations of previous approaches that did not account for non-ideal solution behavior and the complexity of multicomponent evaporation. The accuracy achieved in predicting the original composition, even for highly weathered samples, significantly enhances the potential for linking fire-scene gasoline samples to suspects. This advancement can significantly aid in arson investigations by improving the reliability and effectiveness of forensic evidence analysis, leading to better investigative outcomes.

This study presents a novel methodology for analyzing weathered gasoline in arson investigations. By combining thermodynamic modeling, COSMO-RS calculations for activity coefficients, and artificial neural networks trained on extensive field data, the researchers accurately predict the unweathered composition of gasoline samples with deviations of less than 4%, even for highly weathered samples. This allows for robust forensic comparisons between crime scene and suspect samples, improving the effectiveness of arson investigations. Future research could focus on incorporating other weathering factors and expanding the dataset to further refine and improve the model's predictive accuracy and applicability across different geographical regions and gasoline formulations.

The model currently focuses on evaporation as the primary weathering mechanism, simplifying the complex processes involved. Factors like preferential adsorption, combustion products, microbial degradation, and fire extinguishing agents are not explicitly considered. While the ideal case is a useful starting point, future iterations could incorporate these factors for increased accuracy and real-world applicability. The accuracy of the ANN model relies heavily on the quality and quantity of the training dataset, particularly for highly weathered samples. Expanding the training dataset could further enhance the model's performance in these cases. The COSMO-RS method used for calculating activity coefficients is relatively accurate but could show some limitations in extremely concentrated solutions. More sophisticated or experimentally validated methods might offer improvements for highly weathered samples.