knowledge_base / classified: UNCLASSIFIED

AI RESEARCH

Wiki-style briefing on UK MOD applications of artificial intelligence and machine learning across defence test & evaluation. Compiled from open-source UK public-domain sources.

~/research/overview.md

UK MOD AI Use Cases for Defence Test & Evaluation

A comprehensive research repository covering AI/ML applications across weapons qualification, fragmentation trials, and uncrewed autonomous systems testing (air, surface, subsurface, ground).

Status: Research briefing (UK public-domain sources) Classification: UNCLASSIFIED Compiled: April 2026 Scope: UK MOD internal research reference


Purpose

This repository consolidates open-source research on the application of artificial intelligence and machine learning across the UK Ministry of Defence's test and evaluation (T&E) enterprise. It is organised by domain and designed to support decisions on capability investment, research prioritisation, safety-case development, and engagement with Dstl, DE&S, the Defence AI Centre (DAIC), and NATO partners.

The work is deliberately UK-centric — citing JSP 520, JSP 936, DEFSTAN 00-970 Part 9, RA 1600, the UK Defence AI Strategy, NATO STANAGs, and named UK ranges, programmes and industrial actors — so it can serve as a starting point for MOD-aligned capability planning.


Repository Structure

FolderTopicFocus
01-weapons-qualification-s3/Weapons qualification & Safety and Suitability for Service (S3)JSP 520, DEFSTAN 00-056/07-85, STANAG 4297/4439, DOSG/DOSR, AI for safety case generation, surrogate models, anomaly detection, active learning DoE
02-fragmentation-trials/Fragmentation & arena trialsSTANAG 2636/4496, pit/arena tests, flash X-ray (Scandiflash), PDV, AI for fragment detection, tracking, X-ray denoising, PINNs for blast fields, digital twins
03-uav-testing/Uncrewed Air Vehicle (UAV) testingRA 1600, DEFSTAN 00-970 Part 9, DO-365, ASTM F3442, Protector RG Mk1, GCAP, Project Alvina, AMLAS, sense-and-avoid, swarming, SLAM
04-uncrewed-maritime-ground/USV / UUV / UGV testingProject CABOT, NavyX, Project Wilton, Cetus/Manta XLUUV, Project THESEUS, Robotic Platoon Vehicles, BUTEC, HORIBA MIRA, MASS Code of Practice
05-cross-cutting-themes/Policy, assurance, MLOps, ethicsJSP 936, Defence AI Strategy, AAIP/AMLAS/SACE, synthetic environments & digital twins, MLOps for defence, Article 36, meaningful human control
GLOSSARY.mdAbbreviations & acronymsDefence, regulatory, and ML terminology used across the repo

How to Read This Repository

  • Start with 05-cross-cutting-themes if you are new to UK defence AI governance — it frames every downstream domain chapter and covers JSP 936 and the assurance frameworks (AMLAS/SACE) that constrain the rest.
  • Jump directly to a domain chapter if you are researching a specific programme (e.g. Protector → chapter 03, Cetus XLUUV → chapter 04, IM testing → chapter 01).
  • Each domain chapter follows the same structure: regulatory/programmatic landscape → evidence classes → concrete AI use cases (with maturity/TRL assessment) → UK actors → references.

Key Cross-Cutting Insights

  1. Assurance is lifecycle-wide, not a gate. AMLAS and SACE (University of York) are the defensible UK methodologies for ML in autonomous systems; they replace single-point certification with continuous assurance from scoping to deployment. JSP 936 (Part 1 Directive, November 2024) mandates their principles across the MOD.

  2. Synthetic environments and digital twins are central. Every domain in scope — weapons qualification through to USV/UGV — benefits from high-fidelity physics/sensor simulation (Isaac Sim, AirSim, VBS4, OneSAF, Unreal Engine), with domain randomisation to close sim-to-real gaps.

  3. Live trials remain the regulatory anchor. AI is an accelerator and an efficiency multiplier, not a replacement. Bayesian design-of-experiments and active learning can reduce STANAG 4439 rounds or arena shots by 25–40 %, but DOSR, DOSG, MAA, and NATO standardisation bodies still require live-fire evidence and human sign-off.

  4. Explainability is a compliance requirement, not a nice-to-have. Article 36 reviews, Responsible AI Senior Officer (RAISO) sign-off under JSP 936, and the UK's "meaningful human control" position at the UN CCW all require that autonomous system decisions be auditable post-hoc. SHAP, LIME, attention visualisation, and concept activation vectors all feature.

  5. Data is sovereign. Federated learning and on-premise MLOps stacks (MLflow, DVC, Label Studio / CVAT on classified networks) are how the UK reconciles the Defence Data Framework with AUKUS and NATO collaboration.

  6. The UK ecosystem is federated. Dstl, DAIC, DE&S, DASA, the MAA, MSIAC, the Centre for Assuring Autonomy (York), Cranfield, Turing Institute, BAE, QinetiQ, Leonardo, Thales, MBDA, GA-ASI, plus a growing SME tier (Adarga, Helsing, Mind Foundry, Faculty AI, Ripjar) together form the capability base.


UK MOD Policy Anchors


Caveats

  • All content is drawn from open sources. Classified Dstl, DOSG, MAA, or NATO RESTRICTED material is not included.
  • Maturity / TRL assessments reflect public-domain evidence as of April 2026; internal MOD programmes may be further advanced but unpublished.
  • Use case recommendations and pilot programme proposals are forward-looking and subject to engagement with Dstl, DOSR, the MAA, the DAIC, and the Defence AI and Autonomy Unit.
  • Where content is speculative (e.g. internal MOD digital-twin maturity, unpublished Dstl work), it is flagged in-section.