02 — Fragmentation Trials

AI/ML applications in UK MOD fragmentation characterisation and lethality assessment

---

▌Executive Summary

Fragmentation trials characterise warhead and munitions behaviour under detonation — measuring fragment mass distribution, velocity and spatial distribution to support lethality prediction, casing design and safe-separation distances. Modern trials generate terabytes of high-speed video, flash X-ray radiographs and sensor time-series per detonation. Manual analysis — fragment counting, silhouette interpretation, trajectory reconstruction, panel inspection — is labour-intensive and error-prone, and each arena shot costs £50k–£200k. AI/ML offers substantial near-term gains across detection, tracking, surrogate modelling, digital twins and automated reporting.

---

▌1. Purpose and Methodology

1.1 Objectives of Fragmentation Trials

• Fragment mass distribution (FMD) — binning by weight to distinguish anti-personnel (~1–50 g) vs anti-materiel (>50 g) populations.
• Fragment velocity — kinetic energy and penetration capability.
• Spatial distribution — directional bias, density patterns, dispersion geometry.
• Lethality parameters — effective fragment count, lethal radius, predicted casualty radius.

Feeds directly into:

• Weapon design optimisation (casing geometry, explosive fill, liner materials).
• Safety separation distances for storage and handling.
• Compliance with NATO IM protocols (STANAG 4496, 4589, 2920).

1.2 Standard Test Methods

Pit / Sandbox Tests. Warhead detonated into a closed arena of sawdust, sand or Celotex; fragments recovered, sorted and weighed. Reference: ITOP 4-2-813; modern STANAG 2636.

Arena Fragmentation Tests (AFT). Witness plates (Celotex, woolboard, synthetic materials) at fixed radii (5–20 m). Instrumentation: PHANTOM high-speed cameras (>10,000 fps, 38 ns exposure), recovered panels, screen sensors and flash X-ray.

Flash X-Ray Imaging. Scandiflash systems via Hadland Imaging — 80–1200 kV, 20–35 ns pulses, 1–4 mm focal spot. Multiple flashes reconstruct trajectory inside the expanding fragment cloud.

1.3 UK Test Ranges

1.4 Key Instrumentation

1.5 Lethality Modelling Tools

• AVAL (Thales/QinetiQ) — lethality radius prediction (restricted).
• ORCA / VAULT (US DoD) — collateral damage and vulnerability (restricted).
• MERCAT (Dstl) — UK casualty estimation.
• LS-DYNA, CTH (Sandia), ALE3D (LLNL), AUTODYN — hydrocode simulation.

---

▌2. The Data Problem

A single arena test generates 10⁶+ video frames plus radiographs, sensor arrays and panel imagery. Labour-intensive manual processes dominate:

1. Fragment counting and mass binning (days per test).
2. Frame-by-frame trajectory analysis in high-speed video.
3. Radiograph interpretation for fragment identification.
4. Witness-panel penetration-hole analysis.
5. 50–200-page trial report synthesis.

Core data challenges: fragment overlap/occlusion, motion blur, X-ray noise and streak artefacts, 2D-to-mass inference, sparse sensor coverage, and asymmetric blast waves.

---

▌3. AI/ML Use Cases

3.1 Fragment Detection & Counting (High-Speed Video and Radiographs)

• YOLO family (v5/v8/v9) — real-time bounding-box detection; fast inference for frame-by-frame processing.
• Mask R-CNN — instance segmentation handles overlapping fragments via ROI Align.
• Workflow: annotate 500–2,000 arena frames (CVAT/Roboflow); augment with physics-rendered LS-DYNA synthetic data; link detections across frames via Hungarian-algorithm tracking.

References. MDPI 2021 deep-learning fragment detection for AFT; Faster R-CNN for AFT (Korea Inst.); Mask R-CNN architecture.

3.2 Fragment Mass Estimation from Silhouettes

• Silhouette features (area, perimeter, aspect ratio, circularity, Hu/Zernike moments) via UNet or Mask R-CNN masks.
• Regression: Gaussian Process (probabilistic, with prediction intervals), Random Forest / XGBoost (interpretable), MLP (non-linear).
• Training data from recovered-fragment photograph/weight pairs; augment with LS-DYNA synthetic shapes.
• GPR uncertainty bounds are critical for DOSG compliance reporting.

3.3 Fragment Velocity & Trajectory Reconstruction

• Deep-learning optical flow (RAFT, FlowNet2) for dense velocity fields on noisy imagery.
• Transformer-based trackers (e.g. "Trackformers") — EPJC 2025.
• 3D triangulation using multi-camera arenas; bundle adjustment with ballistic-equation constraints.
• Physics-informed trajectory fitting rejects unphysical solutions.

3.4 Flash X-Ray Image Enhancement

• Denoising: DnCNN, U-Net VAE, GAN-based (DNGAN) for radiography.
• Super-resolution: SRGAN, ResNet/DenseNet upsamples 2–4×.
• Artefact removal: conditional GAN or PDE-based inpainting for streaks.
• Time-series tracking: multi-flash detection + Hungarian assignment; triangulation if multi-angle radiographs available.

References. Deep-learning X-ray denoising (ScienceDirect); HRNet CT denoising.

3.5 Blast-Wave & Pressure-Field Reconstruction from Sparse Sensors

Physics-Informed Neural Networks (PINNs) embed conservation laws (Euler equations, continuity, momentum) as soft constraints:

L = L_data + λ · L_physics

Network maps u_θ(x, t) → P(x, y, z, t) with automatic differentiation of PDE residuals. Supports uncertainty quantification via ensembles and handles Neumann/Dirichlet boundaries naturally. Sensitive to constitutive-model mismatch — validation against unused sensor data is mandatory.

References. PINNs for wave and pressure (ScienceDirect); AGU wave propagation.

3.6 Surrogate Models for Lethality / Collateral Damage Prediction

• GP and neural-network surrogates mapping warhead parameters → lethality metrics (effective fragment count, lethal radius, V50).
• GPR for small datasets (50–500 simulations); neural nets for 10+ parameter spaces.
• Multi-fidelity combining simplified FEA with full hydrocode, plus Bayesian optimisation (Expected Improvement) for design exploration.

3.7 Bayesian DoE & Adaptive Sampling

Arena trials cost ~£50k–£200k per shot. Adaptive sampling strategies:

• Initial 3–5 shots at nominal; fit surrogate; pick next shot to maximise Expected Improvement / UCB / Information Gain.
• Multi-objective (lethality vs mass) and constrained optimisation (diameter, explosive mass).
• Documented 30–50 % trial reduction vs full-factorial DoE.

3.8 Digital Twins & Model Calibration

Paired hydrocode model + real-time/post-trial data feed. Calibration via particle filter, Bayesian MCMC / variational inference or adjoint-gradient optimisation. Observations include recovered FMD, PDV velocities, radiograph features (via CNN feature extraction) and pressure gauges.

References. IEEE digital-twin calibration; Cambridge Core continuous calibration.

3.9 LLM-Assisted Report Generation

RAG pipeline: structured trial metadata + AI-derived quantities + figure templates + STANAG-aware prompts → draft 50–200-page report with compliance statements. Studies in clinical-trial reporting show 50–75 % time reduction. Human review is mandatory — hallucination risk on standards, figures and compliance claims is unacceptable.

3.10 Unsupervised Morphological Clustering

• Structure-from-motion or 2D silhouette features per recovered fragment.
• GMM/Bayesian GMM (with BIC-based cluster count), DBSCAN, spectral clustering.
• Clusters map to lethality sub-types (flat splinter vs spheroidal chunk).

---

▌4. Maturity Summary

---

▌5. Actors

UK

• Dstl (Platform Systems & Weapons Division; MERCAT).
• Cranfield University CDE + COTEC — MSc Guided Weapons, EOE; active warhead and fragmentation research.
• QinetiQ — MOD Pendine, Shoeburyness, Aberporth operator; High Explosive Formulation team; IM research; BROACH warhead partner.
• MBDA — missile/warhead integration (Meteor, Scalp, Storm Shadow).
• Thales UK — sensors, seekers, AVAL.
• BAE Systems Global Combat Systems (Munitions) — warhead development; SPEAR trials.

International

• US ARL, Sandia (CTH, PDV), LLNL (ALE3D, high-speed imaging).
• Imperial College London — blast/shock wave research.
• NATO STO AVT Panel and MSIAC.

---

▌6. Assurance Considerations

Reproducibility and Validation

AI-derived measurements must achieve inter-observer agreement, reference-standard validation (manual vs AI head-to-head on 10–20 arena tests), documented UQ (aleatoric + epistemic), and version-controlled traceability.

DOSG / NATO Approval Pathway

1. PoC on archived AFT data from 3–5 previous trials.
2. Validation trial with parallel manual + AI measurement.
3. Standardisation proposal to NATO STO or UK MOD for STANAG/DefStan update.
4. Operational acceptance.

Indicative timeline: 18–24 months PoC to operational acceptance.

Uncertainty Quantification

Aleatoric via probabilistic outputs; epistemic via ensemble disagreement. Final reports should include confidence intervals e.g. "Fragment count: 847 ± 23 (1σ)".

---

▌7. Recommendations

Short-term (0–12 months).

• Fragment detection PoC on archived AFT video (Dstl + Cranfield COTEC / QinetiQ), £150k–£250k.
• Flash X-ray denoising PoC on archived Scandiflash radiographs, £100k–£150k.

Medium-term (1–2 years).

• Validation trial series with parallel human + AI measurement, £500k–£800k.
• Digital-twin prototype integrating LS-DYNA + Bayesian calibration, £250k–£400k.
• Bayesian DoE retrospective study, £80k–£120k.

Long-term (2–5 years).

• Operational deployment, £200k–£300k/yr ops + development.

Total £1.5M–£2.5M over 5 years positions UK MOD as leader in AI-augmented weapons testing.

---

▌8. Key References

NATO and Standards

• NATO NSO database — STANAG 2636, 4496, 4589
• STANAG 4496 (GlobalSpec)

Published Research

• Deep-learning fragment detection AFT (MDPI 2021)
• ML to predict warhead fragmentation in-flight (Larsen thesis 2022)
• Expanding-warhead fragmentation (Thin-Walled Structures 2014)
• Warhead lethality multi-objective optimisation (IOP 2022)

Instrumentation

• PHANTOM high-speed ballistics
• Scandiflash flash X-ray
• PDV (Tektronix)