AI can rank a geophysical anomaly, warn that a pump is behaving differently or help plan a vehicle route. It cannot confirm an orebody without physical evidence, certify ground stability or make an unguarded autonomous machine safe.
The operating pattern is observe → infer → investigate → authorise → execute → verify. Every model output should show what was measured, what was inferred, the uncertainty and the competent person who owns the next decision.
The source review and regulatory boundaries below are current to 31 July 2026. Mining, quarrying, planning and environmental regulation differ by activity and nation. HSE and the Mines Regulations framework cited below apply in Great Britain; Northern Ireland has separate HSENI and legislative arrangements. Environment Agency permitting guidance cited below applies to England, with separate regulators in Wales, Scotland and Northern Ireland. Confirm whether a site is legally a mine, quarry, borehole, waste operation or another workplace before mapping controls.
Start With the Real UK Operation
The old promise of “autonomous mines” often imports examples from very large overseas surface operations. A UK use case may instead involve an underground mine, a quarry, a proposed critical-mineral project, drilling, mineral processing, a spoil tip or investigation of historic workings.
Define the decision and boundary:
| Use case | AI may support | Evidence it cannot replace | Accountable control |
|---|---|---|---|
| Exploration | Rank targets and integrate geology, geochemistry and geophysics | Sampling, drilling, assays and geological interpretation | Competent exploration programme and licence/land permissions |
| Ground and water risk | Detect movement, pressure or inflow patterns | Geotechnical and hydrogeological assessment | Mine operator’s major-hazard controls and specialist judgment |
| Mobile plant | Detect people/objects and optimise routes | Segregation, braking, guarding and safe systems of work | Traffic plan, operator/controller and emergency stop |
| Process optimisation | Suggest feed, recovery or energy settings | Metallurgical tests, permit limits and plant interlocks | Approved operating envelope and competent control room |
| Predictive maintenance | Prioritise inspection from condition data | Isolation, physical inspection and proof of repair | Maintenance authority and return-to-service process |
| Environmental monitoring | Detect trends in water, dust, vibration or waste | Calibrated compliance monitoring and approved methods | Permit holder, management plan and regulator reporting |
Record the mine or quarry, asset, coordinate system, timestamp, unit, model version, operating mode, uncertainty and authorised user. Similar names or map points are not sufficient identity for a borehole, shaft, seam, bench or vehicle.
Preserve Geological Uncertainty
Geological models interpolate between sparse observations. Machine learning can find relationships across maps, cores, assays and geophysics, but a high target score remains a hypothesis.
Build a provenance register:
- survey, instrument, method, calibration and acquisition date;
- borehole collar, downhole path, datum and coordinate reference system;
- core recovery, sample interval and handling;
- laboratory method, detection limit, quality-control samples and certificate;
- historic-map scale, accuracy and known revisions;
- any compositing, interpolation or synthetic values; and
- licence, permitted reuse and original source.
The British Geological Survey’s BritPits dataset illustrates why metadata matters: it covers surface and underground mineral workings across Great Britain, Northern Ireland, the Isle of Man and Channel Islands, but offers different index and licensed data products. A point showing a working is not a current reserve, safe access route or legal permission.
For coal and historic workings, the Mining Remediation Authority’s July 2026 data-access guide describes records for England, Scotland and Wales, including abandonment plans, mine entries and licence information. The authority warns through its product-specific metadata and licensing routes what each dataset was created to support. Keep source caveats with any model feature.
Split evaluation spatially. Neighbouring samples from the same structure should not appear in both training and test sets. Also test by deposit style, survey campaign and laboratory. Report target precision at a stated investigation budget, missed known mineralisation, uncertainty calibration and the number of recommendations that materially change after new drilling.
Require a geologist to approve a drill target and document alternatives. Stop when coordinates, units or data rights are unclear. A colourful prospectivity map is not a resource estimate.
The government’s Vision 2035 Critical Minerals Strategy, updated in January 2026, sets policy ambitions for resilient supply, domestic production, processing and recycling. It does not lower planning, safety, environmental, financing or technical evidence requirements for a project.
Put Major Hazards Above Optimisation
The HSE’s Mines Regulations 2014 guidance places the principal duty on the mine operator and uses goal-setting requirements focused on major hazards. It highlights competence and suitable rescue provision. An AI supplier is not the mine operator, and a dashboard is not the mine’s health and safety document.
Map every model to the existing control architecture for:
- ground movement, faces, shafts and tips;
- fire, explosion and dangerous atmospheres;
- ventilation, dust and occupational exposure;
- inrushes of water, gas or material;
- electricity and machinery;
- transport and vehicle interaction;
- escape, rescue and communications; and
- abandonment or change of use.
Independent engineered limits must remain authoritative. A ventilation optimiser may suggest fan settings only inside a range approved for the mine’s current state. A movement model may escalate an inspection; it must not dismiss a statutory or geotechnical trigger because the pattern looks normal.
Use conservative missing-data behaviour. A flat gas reading can mean a stable atmosphere or a failed sensor. Compare redundant instruments, validate timestamps and alarm on implausible persistence. Keep local alarms and emergency action available if analytics, communications or cloud services fail.
Quarries are a different legal setting. HSE’s Quarries Regulations 1999 Approved Code of Practice addresses duties to workers and people affected by quarrying. Determine the correct regime rather than copying an underground-mine control set onto a quarry—or treating quarry practice as sufficient underground.
For the project-management interface around excavations and contractors, see our UK construction safety guide.
Bound Autonomous and Remote Plant
Automation should begin in a defined exclusion zone with a machine-checkable operating design domain:
- mapped route, bench, gradient and edge clearance;
- permitted vehicle, load and attachment;
- maximum speed and stopping envelope;
- lighting, dust, rain, fog and surface limits;
- localisation accuracy and loss threshold;
- person, vehicle and obstacle detection coverage;
- communications latency and dead zones;
- interaction with manual plant;
- authorised start, pause, recovery and maintenance; and
- minimum-risk condition.
Physical segregation is stronger than a vision model. Use barriers, controlled access and positive isolation where practicable. Detection should be an additional layer, not permission to mix people and moving machinery without a safe traffic plan.
Test payload, slope, tyre condition, loose material, water, glare, dust, occlusion and sensor contamination. Report detection probability and false/late alerts by range and object class, stopping-distance margin, localisation loss, emergency-stop success and unauthorised-domain entry.
No worker should enter to recover an immobilised vehicle until its state, energy isolation and control authority are known. Log every remote command and handover. Train staff on degraded modes rather than only normal production.
Robots and drones can reduce exposure during mapping or inspection, but their images remain observations. Our UK drone operations guide covers the separate aviation route for aerial work.
Keep People and Competence in the System
Wearables, cameras and proximity tags can support emergency response while also creating continuous worker monitoring. State the safety purpose, minimum data, who sees it, retention and consequences of an alert. Do not repurpose location traces for productivity scoring without a new lawful and fair assessment.
Consult workers and safety representatives before the pilot. Test fit, accessibility, alarm audibility, language, false alarms and whether a person can challenge an incorrect record. A worker should not carry the risk of a dead battery or poor underground coverage that management chose.
Competence includes understanding the model’s limits. Train users to recognise out-of-domain data, sensor failure and automation bias. Define who may accept, override or silence an alert, and audit those actions without discouraging justified intervention.
Control Waste, Water and Environmental Claims
AI may forecast water quality, dust or plant efficiency. Regulatory records still require approved instruments, sampling, methods and competent review.
The Environment Agency’s waste environmental-permit guidance states that an operator in England may need a permit for mining waste and links to separate national guidance. A permit can require management plans, monitoring and records. A model does not change an emission limit or authorise an unlisted waste route.
Maintain source readings, calibration, laboratory certificates, permit mapping, alert review and correction history. Test the model on exceedances, sensor drift, seasonal flow and unusual operating states. Report missed exceedances and time to action, not only average forecast error.
Do not claim that automation makes extraction “sustainable” from energy per tonne alone. Track material recovery, total energy and water, waste by class, land disturbance, emissions, rehabilitation and safety outcomes over the same boundary. Separate an operational improvement from changes in ore grade, production rate or equipment.
Secure Operational Technology
Plant control, ventilation, pumping and access systems are cyber-physical. Inventory assets and connectivity; separate safety-critical networks; restrict supplier and remote access; protect model and configuration updates; monitor commands; and retain tested offline operation.
Replay attacks, wrong coordinates, corrupted maps and compromised credentials belong in the hazard set. A cyber incident can become a vehicle movement, ventilation or water-control incident. Our AI cybersecurity guide covers the broader response lifecycle.
Set Measurable Release Gates
Approve a bounded use case only when every applicable gate passes:
| Gate | Minimum release evidence |
|---|---|
| Legal scope | Mine, quarry, borehole, waste and national regulatory routes confirmed by named owners |
| Data provenance | 100% of critical geological and operational inputs have source, coordinates, units, time and quality status |
| Exploration validity | Spatially separated evaluation passes pre-agreed target precision, miss-rate and uncertainty thresholds |
| Major hazards | Model mapped into the mine/quarry control plan; it cannot suppress an independent statutory or engineered trigger |
| Plant domain | Route, speed, slope, weather, localisation, segregation and payload limits enforced outside the model |
| Detection and stopping | Critical object tests pass detection, latency and stopping-margin thresholds across domain conditions |
| Failure state | Sensor, localisation, communications, model and power failures reach the defined minimum-risk condition every time |
| Workforce | Consultation, competence, alarm, accessibility, privacy and challenge processes completed |
| Environment | Compliance data reconciles to approved instruments and permits; zero missed critical exceedances in validation |
| Operations | Named controller, geologist/specialist, maintenance, incident, rollback and change-control owners active |
Monitor exploration recommendations confirmed or rejected, geological uncertainty, sensor faults, safety-trigger conflicts, domain exits, emergency stops, worker overrides, maintenance findings, permit alerts, missed exceedances and cyber events. Pause when a critical input loses provenance, the site changes beyond the domain, a safety layer is unavailable or performance breaches its floor.
Mining AI is credible when it respects the gap between a pattern and the ground. It should help competent people investigate sooner and operate inside stronger controls—not turn uncertain geology or a safety-critical workplace into a confident animation.
