Emerging Physics: Terahertz, Hyperspectral, Quantum and AI Imaging

Terahertz radiation lies between the microwave (frequencies below ~100 GHz) and the mid-infrared (wavelengths below ~30 micrometres) regions of the electromagnetic spectrum. The approximate THz band for imaging applications runs from 0.1 to 10 THz, corresponding to wavelengths of 30 micrometres to 3 mm. This band was historically described as the "THz gap" because neither the electronic oscillator-based sources used for microwave generation nor the thermal sources used for infrared emission could produce adequate power at THz frequencies.

Generating THz radiation. The dominant laboratory method for forensic-quality THz imaging is time-domain spectroscopy (THz-TDS). A femtosecond pulsed laser (typically Ti:sapphire at 800 nm or fibre-based erbium-doped at 1550 nm) splits into a pump beam and a probe beam. The pump illuminates a photoconductive antenna (a semiconductor biased at a few tens of volts with interdigitated metallic electrodes), generating a sub-picosecond current pulse that radiates a broadband THz pulse. The same THz pulse, after transmission through or reflection from the sample, gates the probe beam through a second photoconductive antenna or electro-optic crystal, allowing time-resolved amplitude measurement across the THz band. The Fourier transform of the time-domain signal yields the amplitude and phase spectrum, from which the sample's refractive index and absorption coefficient can be extracted.

Penetration and contrast mechanisms. THz radiation penetrates non-conducting, non-polar materials (paper, cardboard, clothing fabric, polyethylene packaging) with low attenuation, while being reflected or absorbed by conductive materials (metals, water, ionic solutions, graphite). This penetration profile makes THz imaging qualitatively different from X-ray (which distinguishes materials primarily by density and atomic number) and from optical imaging (which distinguishes materials by their visible-wavelength optical properties). For security screening, the contrast between a concealed metallic object and the surrounding clothing fabric is high at THz frequencies.

Forensic document examination. A forged document may have an original layer of printing partially erased and overprinted with new information, or may have an additional layer adhered onto the original. THz time-domain imaging in reflection geometry can non-destructively image the interfaces between layers: because each layer interface reflects a portion of the THz pulse, the time delays between successive reflected pulses encode the layer thicknesses and the dielectric properties of each layer. Research groups in Germany (Physikalisch-Technische Bundesanstalt), Japan (NICT), and the UK (Teraview Ltd) have demonstrated sub-millimetre depth resolution in layered-document imaging. The technique preserves the original document (no chemical treatment, no disassembly) and is therefore compatible with subsequent conventional document examination.

Concealed-weapon detection. L3 Technologies (now L3Harris) deployed the ProVision ATD (Active Terahertz Detection) system in numerous airports and public-venue security checkpoints from approximately 2013. The European Commission's Airport Security Regulation (EU) 2015/1998 permits the use of body scanners including THz-based systems at EU airports under specific operator-training requirements. The US TSA has deployed millimetre-wave body scanners (operating at 24-30 GHz, technically at the boundary of the THz definition) at major airports since 2009, generating significant debate about privacy (full-body images) that led to the adoption of a generic "chalk figure" avatar display rather than a photorealistic body image. India's Bureau of Civil Aviation Security (BCAS) has piloted body-scanner technology at major airports (Indira Gandhi International Delhi, Chhatrapati Shivaji Maharaj International Mumbai) under a phased implementation plan, using millimetre-wave systems from Rohde and Schwarz and Nuctech.

Pharmaceutical counterfeiting. THz spectroscopy can identify the crystalline polymorph of an active pharmaceutical ingredient (API) through sealed packaging, because different polymorphs have distinct THz absorption spectra. Counterfeit medications that contain the correct API in the wrong polymorph (which may have different bioavailability), or that contain incorrect filler materials, can be distinguished from authentic products without opening the packaging. This application is under active validation at the UK Medicines and Healthcare Products Regulatory Agency (MHRA), the US FDA, and the Indian Central Drugs Standard Control Organisation (CDSCO).

Conventional digital photography captures three broad spectral channels (red, green, blue) at each pixel. Hyperspectral imaging extends this to hundreds of narrow spectral bands, typically covering 380 to 2500 nm (visible through near-infrared) in bands 5-10 nm wide. The result is a three-dimensional data cube: two spatial dimensions and one spectral dimension. For forensic examination, this means that any spectral difference between materials that is too subtle to see with the eye, or that falls outside the visible wavelength range, becomes a quantifiable, spatially resolved signal.

Document examination systems. The Video Spectral Comparator (VSC) platform from Foster and Freeman has been the workhorse hyperspectral document-examination system in forensic laboratories for over two decades. The VSC6000 and VSC8000 models illuminate the document with selectable narrow-band light sources (LEDs, filtered lamps) and capture images at wavelengths from 365 nm (UV) to 1000 nm (near-IR) through a series of bandpass filters. The newer FORAY platform uses a push-broom hyperspectral sensor to acquire a full spectral cube without changing filters, reducing examination time. UK FSS (Forensic Science Service, now CAST, the Centre for Applied Science and Technology) and ENFSI document-examination working groups have validated VSC-family systems for distinguishing ink types, detecting erasures and additions, and sequencing ink strokes.

Ink-stroke sequencing. When two ink strokes cross, determining which was applied first is a classic forensic document question with significant legal consequences (establishing the chronology of a will codicil or a contract amendment, for instance). Hyperspectral imaging can sometimes distinguish the crossing order based on the penetration depth and spectral character of the inks at the intersection, particularly when the two inks have different spectral profiles. The technique does not work for all ink pairs: inks of similar composition may show insufficient spectral contrast. The ENFSI Document Examination Working Group's guidelines on ink examination specify that hyperspectral ink-sequencing evidence should be accompanied by a statement of the spectral contrast between the two inks and the uncertainty of the determination.

Bloodstain ageing. Fresh bloodstains are oxygenated haemoglobin (HbO2), with characteristic absorption peaks at approximately 415, 541, and 577 nm. As blood desiccates, haemoglobin converts to methaemoglobin (Met-Hb), then to haemichrome and haematin over hours to days. Each of these species has a distinct reflectance spectrum in the visible and near-infrared, and hyperspectral imaging of bloodstains can track this spectral evolution. Research from the Netherlands Forensic Institute (NFI), the German Bundeskriminalamt (BKA), and academic groups at the University of Amsterdam and University of Strathclyde have proposed bloodstain-age estimation models based on the ratio of HbO2 to Met-Hb derived from hyperspectral imaging. As of 2024, these models are research-grade: they show promise in controlled laboratory conditions but have not been validated for routine casework because the conversion rate depends on temperature, humidity, substrate, and exposure to light, all of which vary unpredictably at crime scenes. The NFI has piloted the technology in Dutch criminal cases with appropriate uncertainty statements.

Wound documentation and bruise enhancement. Hyperspectral imaging in the visible and near-infrared can enhance the contrast of subsurface haemorrhage (bruising) that is not visible to the naked eye, because haemoglobin absorbs strongly at 415 nm and 540 nm while melanin absorption is broad and relatively featureless. This differential absorption allows bruises to be distinguished from normal skin pigmentation variation using spectral unmixing algorithms. UK forensic pathologists and clinical forensic medicine practitioners have used UV reflectance imaging (a single-band subset of the hyperspectral approach) for bruise documentation for several decades (the Vogeley 2002 UV bruise protocol); full hyperspectral imaging provides a richer dataset.

Quantum sensing uses quantum mechanical phenomena to achieve sensitivity levels that are fundamentally inaccessible to classical instruments. The phrase covers several distinct technologies at different stages of forensic application.

Nitrogen-vacancy (NV) centres in diamond. A nitrogen-vacancy centre in diamond is a point defect where a nitrogen atom sits adjacent to a lattice vacancy. The NV centre has electron-spin energy levels that can be initialised, manipulated, and read out using laser light and microwave fields at room temperature, a unique combination that makes NV-based magnetometry practical outside of cryogenic environments. Because the NV centre's resonance frequency shifts with applied magnetic field (Zeeman effect), measuring the spin resonance allows direct measurement of the local magnetic field with spatial resolution determined by the distance from the NV to the sample surface, potentially as fine as a few nanometres. Research groups at MIT, Harvard, and Stuttgart have demonstrated that NV magnetometry can image the magnetic field of a single neuron's action potential and map the magnetic signature of iron nanoparticles at the single-particle level. Forensic applications under investigation include: mapping the residual magnetisation in partially-fired cartridge cases (the pattern of residual magnetic field in the case body encodes the peak pressure and temperature during firing), imaging magnetic signatures of latent fingerprints deposited on magnetic substrates, and detecting biomagnetic signals from decomposing tissue samples.

SQUID magnetometry. Superconducting quantum interference devices (SQUIDs) are ultra-sensitive magnetic-field detectors that exploit quantum flux quantisation in superconducting loops. SQUIDs require cryogenic cooling (typically liquid helium or liquid nitrogen for high-temperature superconducting variants) and are not portable. They have been used in forensic contexts for decades in the specific application of detecting ferromagnetic residues in post-blast debris: the magnetic susceptibility signal from iron-containing explosive residues can be detected by SQUID magnetometry at concentrations below what energy-dispersive X-ray spectroscopy can resolve. The US Army Research Laboratory and UK Defence Science and Technology Laboratory (Dstl) have both investigated SQUID-based explosive-residue detection.

Portable NMR. Conventional nuclear magnetic resonance spectroscopy requires a large superconducting magnet (typically 7-21 Tesla, weighing hundreds of kilograms, requiring liquid helium cooling). The development of permanent-magnet benchtop NMR instruments (such as the Magritek Spinsolve, operating at 1-2 Tesla) and, more recently, portable low-field NMR systems (0.05-0.5 Tesla, battery-operated) has made NMR accessible outside the central laboratory. Forensic applications include: non-destructive identification of controlled substances in sealed packaging (important for clandestine-laboratory investigation where destructive sampling would compromise the scene), quality authentication of petroleum products in fuel-fraud investigations, and water-content mapping in foodstuffs for adulteration detection. The CDSCO in India and the MHRA in the UK have both conducted feasibility studies on portable NMR for pharmaceutical authentication.

Validation status. All three quantum sensing approaches are at research or early-pilot stage for forensic casework. None has a published OSAC or ENFSI validation study as of 2024, and none is referenced in a court-admissibility ruling as an authenticated forensic method. The path from demonstrator to casework-validated tool requires: published accuracy and precision data against certified reference materials, proficiency testing across multiple laboratories, blind-test performance evaluation, and a traceable chain of custody for the measurement data. Organisations including the National Physical Laboratory (NPL) in the UK and NIST in the US are involved in developing traceable reference standards for quantum-sensor calibration.

Artificial intelligence, specifically deep learning based on convolutional neural networks (CNNs), has been applied to forensic image-analysis problems including fingerprint matching, footwear-impression comparison, ballistic-mark comparison, face recognition, document-forgery detection, and bloodstain-pattern classification. In each case, the basic framework is the same: a neural network is trained on a large labelled dataset of forensic images, adjusting its internal parameters to minimise classification error on the training set, then evaluated on a held-out test set before deployment.

Fingerprint and footwear impression matching. The FBI's NGI (Next Generation Identification) system uses AFIS (Automated Fingerprint Identification System) algorithms that incorporate machine-learning components for feature extraction and match scoring. Commercial systems including Cogent, Aware, and Idemia's MorphoTrak use proprietary deep-learning architectures. These systems have published False Match Rates (FMR) and False Non-Match Rates (FNMR) characterised by large-scale blind-test evaluations coordinated by NIST's FRVT (Face Recognition Vendor Testing) and FpVTE (Fingerprint Vendor Technology Evaluation) programmes. For footwear-impression matching, the FBI and several European forensic institutes have piloted CNN-based systems; the ENFSI Marks Working Group has published initial guidelines for evaluating AI-assisted footwear comparison tools.

Ballistic mark matching. The National Institute of Justice (NIJ) has funded multiple research projects applying CNN-based matching to SEM images of cartridge-case breech-face impressions and firing-pin impressions. The Congruent Matching Cells (CMC) metric, developed at NIST, is a geometry-based statistical approach that was developed in parallel with deep learning approaches; the two approaches are now being compared in the NIST CSAFE (Center for Statistics and Applications in Forensic Evidence) research programme. No AI-based ballistic system has yet received OSAC (Organisation of Scientific Area Committees for Forensic Science) approval for casework use.

Bloodstain pattern analysis. CNN-based classifiers for bloodstain pattern types (passive drip, spatter, transfer, void, arterial) have been trained on the Pattern Analysis Laboratory at Texas A&M's reference dataset and on datasets curated by the ENFSI Bloodstain Pattern Analysis Working Group. Published classification accuracy on held-out test sets ranges from 85% to 95% for binary tasks (spatter vs non-spatter), with lower accuracy on multi-class tasks. The limitation is that training datasets are laboratory-created under controlled conditions; real-scene bloodstains occur on varied substrates, at varied environmental conditions, and with unknown mixing histories that may not be represented in the training data.

The ENFSI 2023 AI Position Paper. ENFSI (European Network of Forensic Science Institutes) published its AI Position Paper in 2023, covering the use of AI and machine-learning systems in forensic evidence examination across all ENFSI member laboratories. The paper identifies three categories of AI use: operational AI tools (already deployed in casework, e.g., AFIS), emerging AI tools (piloted in casework with validation underway, e.g., CNN-based footwear), and research AI tools (not yet ready for casework). The paper specifies that any AI tool used in casework must have: published performance data from a blind test on a representative case-relevant dataset; a documented uncertainty estimate for each case output; a method for the examiner to verify or override the AI result; and transparency sufficient for court disclosure. ENFSI member laboratories are required to apply the ISO/IEC 17025 quality framework to AI tool validation.

The US framework for validating emerging forensic methods has two primary institutional actors: NIST (National Institute of Standards and Technology) and OSAC (Organisation of Scientific Area Committees for Forensic Science).

NIST's role. NIST's forensic science programme includes the Forensic Science International (NIST-funded research), the OSAC secretariat function, and direct metrology work on calibration and reference materials. For AI systems specifically, the NIST AI Risk Management Framework (AI RMF, published January 2023) provides a voluntary guidance document for organisations developing or deploying AI across all sectors. The forensic-science extension of the AI RMF, coordinated through OSAC's Digital Evidence and Multimedia Subcommittee, applies the four functions of the AI RMF (Govern, Map, Measure, Manage) to AI tools in forensic evidence processing. NIST's FRVT (Face Recognition Vendor Testing) programme is the most mature application: it has evaluated more than 450 face-recognition algorithms since 2000, producing detailed accuracy data that courts and government agencies use to assess specific system deployments.

OSAC's requirements. OSAC is organised into scientific area committees covering different forensic disciplines, each of which publishes standard methods and requirements documents. A standard or requirement published by OSAC and subsequently listed on the OSAC Registry of Approved Standards is the de facto accreditation reference for US forensic laboratories under the ASCLD-LAB and ANAB accreditation frameworks. For AI-assisted forensic tools, OSAC's relevant subcommittees (Digital Evidence, Marks, Biology) have published or are developing requirements that specify: minimum blind-test dataset size; required performance metrics (FMR, FNMR, likelihood-ratio calibration); mandatory human review before casework output; and disclosure requirements for court proceedings.

The UK Forensic Science Regulator. The UK FSR's Codes of Practice and Conduct (published under the Forensic Science Regulator Act 2021) apply to all forensic science providers operating in the criminal-justice system in England and Wales. The FSR has issued a specific guidance note on AI in forensic science (FSR-G-242, 2023) that aligns with the ENFSI position paper. Under the FSR framework, a forensic provider that deploys an AI tool for casework must register it as a forensic process, validate it against ISO/IEC 17025, and maintain a validation record accessible for court disclosure. The FSR guidance explicitly states that the examiner, not the AI system, is the expert witness, and that the examiner must be able to explain the basis of the AI-assisted conclusion to the court.

India's framework. India does not yet have an equivalent to OSAC or the UK FSR for AI-specific forensic validation. The CFSL (Central Forensic Science Laboratory) and state FSLs apply ISO 17025 accreditation through NABL (National Accreditation Board for Calibration and Laboratories), but the NABL accreditation framework does not currently include AI-specific validation criteria. The Bureau of Indian Standards (BIS) has published guidelines on AI quality management under IS 16676 (aligned with ISO/IEC 42001, AI management systems), and the Ministry of Electronics and Information Technology's (MeitY) AI Policy 2023 references responsible AI principles. However, as of 2025 no Indian regulatory body has issued a specific framework for validation of AI systems in forensic casework, and courts admit AI-assisted forensic evidence under BSA 2023 § 45 expert evidence provisions without a prescribed validation standard. This gap represents a significant area for regulatory development, and several CFSL directors have publicly called for OSAC-equivalent guidance.

The ENFSI and European Commission AI Act. The EU Artificial Intelligence Act (entered into force August 2024) classifies AI systems used in law enforcement, including criminal investigation, as high-risk AI systems subject to mandatory conformity assessment before deployment. This includes AI systems used for forensic evidence analysis. EU forensic laboratories deploying AI-based fingerprint, footwear, or ballistic systems will need to document conformity with the AI Act's requirements, including risk assessment, technical documentation, and logging of AI outputs in casework. ENFSI's 2023 AI position paper was developed partly in anticipation of the AI Act's requirements.

Emerging forensic technologies face a common admissibility challenge: courts admit evidence based on established methodology and peer-reviewed validation, but emerging technologies, by definition, have not yet accumulated the validation record that established methods carry. The solution is not to wait until a technique is no longer "emerging" but to characterise what the existing validation record supports and what it does not.

US Daubert standard. Federal Rule of Evidence 702 and the Daubert framework require that an expert's methodology be based on sufficient facts, employ reliable methods, and reliably apply those methods. For AI-assisted forensic tools, the Daubert inquiry focuses on: does the algorithm have a known and acceptable error rate (from blind-test evaluations like NIST FpVTE)? Has it been subjected to peer review and publication? Is it generally accepted by the relevant scientific community? AI fingerprint systems have extensive NIST evaluation records and have been admitted; CNN-based footwear-comparison systems have more limited records and have faced challenges. Courts in Arizona, California, and New York have admitted AI-assisted fingerprint comparison evidence while requiring disclosure of the system's NIST FpVTE performance data.

UK CPS and FSR standards. Crown Court admissibility of forensic evidence is assessed under Part 19 CPR. The FSR's 2023 AI guidance note explicitly addresses the Daubert-analogous questions: the expert must be able to explain the validation basis, state the error rate, and confirm that the AI output has been verified by human review. No AI-assisted forensic tool has yet been the subject of a published Court of Appeal judgment specifically on its admissibility, but the trajectory of case law from R v. Atkins (facial mapping) and the AFIS fingerprint admissibility record suggest that courts will admit AI-assisted tools with appropriate expert qualification and uncertainty disclosure.

Canadian framework. Canada's Mohan admissibility test (R v. Mohan [1994]) assesses reliability, necessity, absence of exclusionary rule, and expert qualification. The more recent R v. Abbey (2009, Ontario Court of Appeal) added a "gatekeeper" function for the trial judge. Canadian courts have admitted AFIS fingerprint evidence extensively; AI-assisted footwear and ballistic tools are at the research-to-pilot transition and have not yet been the subject of a published Canadian appellate decision.

Australia and ANZFSS. The Australian and New Zealand Forensic Science Society (ANZFSS) has established a Technology Advisory Panel that monitors emerging technologies including AI-assisted analysis. Australian courts admit forensic evidence under the Uniform Evidence Acts using an expert opinion framework; AI-assisted tools would be admitted if the expert can testify to their reliability. ANZFSS is developing a technology-readiness assessment framework aligned with OSAC requirements.

THz in document forgery, Netherlands 2018. The Netherlands Forensic Institute (NFI) applied THz time-domain imaging to a disputed will in a civil inheritance case. The document was claimed to have been altered by overprinting. THz reflection imaging revealed a subsurface layer with different dielectric properties at several character positions, consistent with an additional ink layer deposited over the original printing. The THz data was presented as supporting evidence alongside infrared reflectography and conventional VSC analysis, and the court accepted the combined technical conclusion that the document had been altered.

Hyperspectral bloodstain ageing, pilot study in the UK, 2021. The Forensic Science Service successor bodies (CAST and Dstl) conducted a blind pilot study applying hyperspectral HbO2/Met-Hb ratio analysis to bloodstains of known age deposited on cotton fabric. The study found that HbO2/Met-Hb ratios provided meaningful age discrimination for the first 24 hours but showed overlapping confidence intervals between 24 and 72 hours, with significant substrate dependence. The results were published in Forensic Science International (2022) with the conclusion that the technique requires environmental control data (temperature, humidity, UV exposure) before casework deployment.

AFIS AI in Indian criminal investigation. NCRB (National Crime Records Bureau) operates the National Automated Fingerprint Identification System (NAFIS) deployed across India since 2014 under the Crime and Criminal Tracking Network and Systems (CCTNS) programme. NAFIS uses fingerprint-matching algorithms validated by NCRB to ISO 17025 standards, and its match reports are submitted as expert evidence in criminal trials under BSA 2023 § 45. The system processed over 4 million fingerprint queries in 2023-24 (NCRB annual report), and its outputs have been admitted in numerous sessions courts.

ShotSpotter and AI classification in Chicago, 2021-2024. ShotSpotter's AI acoustic-classification pipeline, described in Section 5 of the Forensic Acoustics topic, has been subject to litigation in Chicago over whether the AI classification process is adequately disclosed to defence counsel. The Circuit Court of Cook County issued a discovery order in 2022 requiring ShotSpotter to disclose its classifier's training dataset, validation metrics, and the specific features extracted from the audio in the case at issue. This discovery order represents the most detailed US court examination of AI forensic tool documentation requirements and is now referenced in defence challenges to AI-based forensic evidence in other jurisdictions.