Predictive modeling of post-blast TNT residue recovery using semi-quantitative LC-MS/MS data and machine learning

Post-blast explosive residue analysis is crucial for forensic investigations, yet evidentiary value is often limited by sampling efficiency rather than instrumental sensitivity. This study investigated the recovery behavior of 2,4,6-trinitrotoluene (TNT) under field-realistic detonation conditions to improve sampling strategies through data-driven modeling. Controlled TNT detonations were conducted at a secured site with bomb disposal experts and crime scene investigators. Residues were collected via swab sampling from metal, wood, and polyester surfaces placed 1-3 m from the blast center. Swabs were pre-wetted with methanol, isopropyl alcohol, or acetonitrile, and samples were analyzed using semi-quantitative LC-MS/MS. The study evaluated how surface material, solvent choice, and distance influence recovery using statistical comparisons and supervised machine learning. Results revealed a clear dependence on surface material; metal yielded stronger, more consistent TNT signals compared to the lower and more variable recovery from wood and polyester. Acetonitrile generally showed higher extraction efficiency, though performance varied by surface. As expected, recoverable TNT decreased significantly with distance. A Support Vector Regression (SVR) model successfully reproduced these experimental trends, providing stable predictions that account for non-linear relationships between variables. These findings demonstrate that field conditions are decisive in residue recovery. Integrating experimental data with machine learning offers a practical method for estimating relative TNT recovery, providing evidence-based guidance for sampling decisions during time-sensitive post-blast forensic investigations.