Emerging Methods: Handheld Raman, LIBS, CT Imaging, ML

Raman spectroscopy measures the inelastic scattering of monochromatic laser light by molecular bonds. The resulting Raman spectrum is a molecular fingerprint: each compound produces a characteristic pattern of Raman shift peaks at specific wavenumber positions. For bulk identification of explosives, ignitable liquids, and drug precursors at checkpoint and scene contexts, Raman spectroscopy became practically viable for field use when miniaturised 785 nm diode lasers achieved the stability and power output needed for library-quality spectra in a handheld form factor.

The primary commercial platforms in field use as of 2024 are the Thermo Fisher TruNarc (optimised for narcotics but used for explosive precursors), the Rigaku Progeny (forensic and security focus), the B+W Tek NanoRam (broad chemical ID), and the Metrohm Mira DS (SERS-enhanced for trace analysis). In the explosives field specifically, the ATF and TSA use Raman instruments for checkpoint screening of suspected explosive materials seized at transportation nodes. The Met Police Counter Terrorism Command uses handheld Raman instruments for in-situ identification of suspect packages.

The limitation in fire investigation is fluorescence interference. Many fire debris matrices (burned carpet, charred wood, sooty surfaces) contain polycyclic aromatic hydrocarbons (PAHs) that fluoresce under near-infrared laser excitation, generating a broad spectral background that can obscure the Raman peaks of the target compound. Modern instruments address this with fluorescence-rejection algorithms (shifted excitation Raman difference spectroscopy, SERDS) or by using a 1064 nm excitation laser (less fluorescence induction at longer wavelengths), which is available on instruments including the Bruker Bravo and the Metrohm Mira DS-1064 variant. Performance on heavily charred or wet debris samples remains instrument-specific, and no ASTM standard for field Raman in fire debris investigation has been published as of 2025.

In India, the DRDO's Instruments Research and Development Establishment (IRDE) in Dehradun has developed indigenous handheld Raman prototypes for defence and security applications. The CFSL and DFSS have evaluated commercial instruments from Rigaku and Thermo Fisher for use in the field; formal integration into NABL-accredited scope remains pending formal method validation. The FSL Maharashtra has trialled handheld Raman for explosives precursor screening following the Mumbai 2011 bombing investigations.

Laser-induced breakdown spectroscopy (LIBS) focuses a high-power pulsed laser (typically Nd:YAG at 1064 nm, with pulse energies of 30 to 300 mJ and pulse durations of nanoseconds) onto the surface of a sample. The laser energy ablates and ionises a small amount of material, producing a short-lived plasma at the focal point. As the plasma cools, the excited atoms and ions emit light at wavelengths characteristic of their elemental identity. A spectrometer integrated with the LIBS instrument captures this emission spectrum, from which elemental composition is determined by peak wavelength and intensity.

For fire and explosives scene investigation, LIBS offers two key capabilities that no other portable technique matches. First, it provides semi-quantitative elemental analysis of bulk materials in situ: metal conductor surfaces can be examined for copper or aluminium, soil samples at the blast seat can be screened for nitrogen and chlorine residues indicative of inorganic explosive components, and paint chips from a fire scene can be characterised for lead, barium, and titanium content that identifies the paint generation and manufacturer. Second, LIBS is entirely non-contact: the laser fires across an air gap of up to several centimetres, and the sample requires no preparation. This is critically important for post-blast investigation where fragile evidence (wiring fragments, container residues, clock mechanism components) must not be disturbed for confirmatory laboratory examination.

Commercial LIBS instruments in field deployment include the SciAps Z-series (used by US military EOD units for explosives-component ID), the Bruker EOS 500 (environmental and security applications), and the LTB Lasertechnik BIGSKY (higher-power research-grade field instrument). The US Army Research Laboratory and the ATF have collaborated on LIBS method development for post-blast inorganic residue identification; several publications have demonstrated reliable detection of ammonium, nitrate, perchlorate, and chlorate residues in soil and on metal surfaces at concentrations relevant to post-blast scenes.

Limitations of LIBS for F/A/E applications include matrix effects (the elemental spectral intensity depends on the matrix, which varies across scene substrates), difficulty with organic compound identification (LIBS provides atomic spectra, not molecular spectra, so organic explosives like TATP or RDX cannot be directly identified by LIBS without complementary Raman or MS analysis), and the high laser energy that can modify the sample surface, potentially compromising subsequent laboratory analysis if LIBS is applied to a sample that will also undergo LC-MS. The recommended workflow, endorsed by the ENFSI EWG discussion paper on in-situ analysis (2022), is LIBS as a first-pass screen followed by confirmatory laboratory analysis on the same or an adjacent sample.

Computed tomography (CT) imaging uses a rotating X-ray source and detector array to produce cross-sectional images (tomographic slices) that are computationally reconstructed into a three-dimensional volumetric dataset. In medical CT, the object being imaged is a patient; in forensic scene CT, the objects range from a single evidence can (investigating its contents and headspace before opening) to large evidence items (assembled IED components submitted for explosive render-safe review before handling) to, in emerging applications, entire rooms or structures.

The application of CT to sealed evidence cans before opening is already in routine use at several accredited laboratories. Opening an evidence can irreversibly changes the headspace composition; CT scanning before opening allows characterisation of the can's internal structure, moisture level, volume of debris, and presence of any metallic fragments or device components before any headspace is disturbed. The CFSL Hyderabad and the ATF Forensic Science Laboratory have both evaluated evidence-can CT protocols. A sealed can containing a suspected IED component or an assembled device can be characterised by CT before explosive-ordnance disposal render-safe, allowing the investigator to understand the device architecture before physical intervention.

For full-scene imaging, the technology transition has been enabled by terrestrial laser scanning (TLS, also called 3D LIDAR scanning) and, more recently, by photogrammetric structure-from-motion (SfM) processing of drone or camera imagery. TLS instruments (Leica BLK360, Faro Focus 3D, Trimble X7) capture millions of point measurements per scan, generating a millimetre-accurate point cloud of the scene. Multiple scan positions are registered together into a unified coordinate system. The result is a 3D model of the scene that survives the scene's physical demolition or remediation, allows virtual re-walk by investigators and experts who were not physically present, and permits precise dimensional measurements years after the scene is cleared.

The UK Fire Investigation Team at BRE has used TLS scene capture on major structural fire investigations including the Grenfell Tower inquiry, where digital 3D models of the building's cladding facade provided the inquiry with spatially precise documentation that outlasted the building's partial demolition. The California Department of Forestry and Fire Protection (Cal Fire) uses drone-mounted LIDAR and SfM photogrammetry for wildfire scene documentation, producing 3D terrain models that locate ignition-point clusters in relation to structures, roads, and known ignition sources.

In India, terrestrial laser scanning has been used by the CSIR-CRRI (Central Road Research Institute) and by engineering consultants for structural-failure investigations, but its adoption in fire investigation by the CFSL or state FSLs has been limited by equipment cost and the absence of formal training programmes. NABL does not yet specify 3D scene documentation requirements in T-126, though the method aligns naturally with ISO/IEC 17025 Clause 7.8.3's requirement for technical records that are reproducible.

Fire debris GC-MS analysis under ASTM E1618 requires the analyst to classify a chromatogram against a set of defined compound classes (gasoline range light petroleum distillates, kerosene range medium petroleum distillates, diesel range heavy petroleum distillates, isoparaffinic products, aromatic products, naphthenic-paraffinic products, and oxygenated solvents). This classification is based on pattern recognition: the relative intensity and position of peaks in the total ion chromatogram, specific extracted ion chromatograms at m/z 55, 57, 71, 83, 91, 105, 119, and 128 (among others), and comparison against the ASTM E1618 Appendix class profiles and reference chromatograms.

This pattern-recognition task is well-suited to machine-learning approaches. The OSAC Fire Investigation and Explosives Subcommittee (F&E) sponsored a validation study in 2022-2023 in which a supervised machine-learning classifier, trained on a dataset of several hundred reference fire debris chromatograms with known classification labels, was evaluated against a test set of novel chromatograms from participating laboratories. The classifier (a random forest architecture trained on extracted ion chromatogram peak ratios as feature vectors) achieved agreement with expert analyst consensus classification at a rate that was statistically non-inferior to inter-analyst agreement among human examiners. Results were published through OSAC and referenced in the 2024 ASTM E1618 revision commentary.

The ENFSI EWG AI working group, active since 2021, has addressed the broader question of AI and machine learning in fire and explosives analysis, including the specific issue of model explainability: how does an analyst or court understand why a machine-learning classifier produced a particular classification? Random forest and gradient boosting models can provide feature-importance rankings (which ion ratios most influenced the classification decision), but this is not equivalent to the line-by-line analytic narrative that a trained examiner can provide. The working group's 2023 position paper recommended that ML tools in accredited F/A/E laboratories be used as decision-support tools (providing a ranked probability for each ASTM E1618 class) rather than as autonomous classifiers, with the examiner retaining final classification authority and the obligation to document their interpretation.

Published research from John DeHaan (Fire-Ex Forensics, formerly California Criminalistics Institute) and the Technische Universiteit Delft (forensic chemistry group) has demonstrated convolutional neural network (CNN) approaches for fire debris chromatogram analysis that achieve comparable performance to random forest models on well-curated training sets, with potentially better generalisation to novel matrices. The challenge in both cases is training data: building a chromatogram dataset that covers the full range of ignitable liquid classes, substrate types (carpet, wood, soil, concrete), weathering stages, and concentration levels requires systematic data collection from multiple accredited laboratories across multiple jurisdictions, which is the current bottleneck for model generalisation.

1. Feature extraction from GC-MS data
   The raw chromatogram is pre-processed: baseline correction, retention-time alignment against a reference standard, and extraction of peak areas at the target m/z ions defined by ASTM E1618. These peak areas (or their ratios) form the feature vector input to the ML classifier.
2. Classification against trained model
   The feature vector is passed to the trained ML model (random forest, gradient boosting, or CNN). The model returns a probability distribution over the ASTM E1618 classification classes. High probability on one class indicates strong model confidence; a spread across multiple classes indicates ambiguity.
3. Analyst review and override
   The examiner reviews the ML output alongside the full chromatogram. If the ML classification is consistent with the examiner's independent interpretation, it is recorded as supporting the conclusion. If the examiner disagrees, the examiner's documented reasoning takes precedence. The ML output is recorded in the case file as a decision-support data point.
4. Uncertainty and interferent documentation
   The analyst documents whether substrate pyrolysis products were present, whether they could have influenced the ML classification, and whether the substrate comparison sample chromatogram was factored into the final determination. This documentation step is mandatory under ISO/IEC 17025 Clause 7.8 and the OSAC F&E guidance.
5. Report generation
   The final classification is stated in the report using ASTM E1618 language. The report notes that decision-support ML analysis was used, identifies the model version and training set vintage, and states that the final determination was made by the named examiner. This satisfies the ANAB supplemental requirement for analyst responsibility.

Wildfire arson investigation presents a scale problem unique among fire investigation subspecialities. The fire scene is measured in hectares or square kilometres, the ignition area is typically a sub-hectare zone that may have been obliterated by the subsequent fire, and the investigative timeline is compressed by landowner pressure to remediate and replant. Ground-based scene examination following conventional NFPA 921 methodology remains the standard for the area of origin determination, but aerial surveillance has become a critical complement for two specific tasks: fire-spread mapping to reconstruct the direction-of-travel back to the origin area, and ignition-point cluster analysis across multiple suspected fires attributed to the same series.

Drone-mounted multispectral cameras capture reflected light in wavelength bands beyond the visible spectrum, typically including near-infrared (NIR, 700-900 nm), red-edge (700-730 nm), and, in more advanced payloads, short-wave infrared (SWIR, 1000-2500 nm). Char depth and surface moisture respond differently across these spectral bands than healthy or recently burned vegetation. Multispectral analysis can distinguish the deeply charred ignition-area ground surface from the lighter char of the fire's perimeter zones, can detect residual heat signatures in the hours after ignition (relevant for arson investigations where the scene is examined shortly after the fire is suppressed), and can map the progression of char intensity across the burned area in a spatial dataset that translates directly to a fire-spread model.

The California Department of Forestry and Fire Protection (Cal Fire) has integrated drone aerial surveys into major wildfire investigations since approximately 2018. The 2018 Camp Fire, which killed 85 people in Paradise, California and became the deadliest wildfire in California history, involved a post-fire drone aerial survey that contributed to the investigation's conclusion about the PG&E electrical transmission line as the ignition source, by mapping the spatial relationship between the area of maximum char depth and the transmission-line corridor. The Australian Forest Fire Management Victoria (AFAC) has used similar aerial methodologies in post-bushfire investigations following the 2019-2020 Black Summer fires, where drone LIDAR combined with multispectral imaging provided terrain-corrected fire-spread maps that ground investigators could not produce within the available time before remediation.

In India, the State Disaster Response Force (SDRF) units in Uttarakhand, Himachal Pradesh, and Nagaland have acquired drone platforms for general disaster response including forest fire monitoring. Forensic application of drone aerial imagery to wildfire ignition-point determination remains at an early stage: the CFSL and state FSLs do not have standardised protocols for drone aerial fire investigation, and NABL T-126 does not address aerial photography as a fire scene documentation method. Research collaboration between the Indian Institute of Remote Sensing (IIRS) in Dehradun and state forest departments has produced multispectral burn-severity mapping for post-fire ecological assessment, but this is ecologically rather than forensically framed.

The gap between a technology's analytical performance and its courtroom admissibility is measured in validation studies, published error rates, and accreditation scope. For all five technologies surveyed in this topic, the performance characteristics are increasingly well-documented in peer-reviewed literature, but the accreditation-scope integration is lagging.

For handheld Raman applied to explosives identification, ANAB and UKAS have accepted laboratory-validated Raman methods within their forensic accreditation programmes, but only where the laboratory can demonstrate an instrument-specific validation against the target analytes at relevant concentrations and matrices. The absence of a shared ASTM or ISO standard for field Raman in F/A/E contexts means that each laboratory must independently validate, creating duplicated effort and preventing inter-laboratory comparison of performance claims. The OSAC Explosives Subcommittee has a working group drafting a technical report on handheld Raman validation criteria for explosive identification (target publication: 2025-2026).

For ML-assisted GC-MS classification, the OSAC F&E 2023 validation study provides the first systematic error-rate data for this application. A random forest classifier applied to the test set produced a misclassification rate (false ASTM E1618 class assignment) of approximately 6 per cent across all classes, with the highest error rates at the boundary between kerosene-range and naphthenic-paraffinic product classes (which are spectrally similar). This 6 per cent inter-class error rate, while higher than expert analyst consensus agreement on the same test set, is the first documented figure available for Daubert error-rate purposes. The ENFSI EWG position paper (2023) recommends that any laboratory deploying ML classification tools report the classifier version, training set vintage, and this error-rate figure in the case report.

For CT scene imaging, 3D LIDAR, and drone multispectral imaging, admissibility challenges have been minimal in practice, because these technologies are treated as documentation methods (generating photographic and three-dimensional records of observable scene conditions) rather than as forensic analytical methods that generate new scientific conclusions. Courts have been consistently receptive to photogrammetric scene models as demonstrative exhibits, applying the same standards that apply to conventional photography: the exhibit must fairly and accurately represent what it purports to represent, and the person who created it must be able to testify to that. The analytical interpretive step (using the multispectral data to identify the ignition point) carries more evidentiary weight and requires the additional foundation of expert testimony from someone qualified in remote sensing and fire investigation.