
AI for Safety
Research theme 02 · Led by Dr Zhibao Mian
In short
This theme turns the previous one around. Instead of asking how to make AI safe, it asks what AI can do for safety engineering, where analysis is still mostly manual, slow, and limited by how much one engineer can keep in mind.
Why it matters
Safety analysis in industry is imagination-driven. An engineer identifies hazards, reasons about what could cause them, and argues that the design reduces risk far enough. The techniques are well established — fault tree analysis, FMEA — and the quality of the result depends heavily on the experience of the person doing it.
That doesn’t scale. Modern systems have more components, more software, more interaction between parts and more configurations than a manual process can work through. Analyses go out of date as soon as the design changes, and the parts most likely to be missed are the interactions nobody thought to look for.
AI suits exactly that gap: searching large spaces, spotting patterns across operating data, and proposing possible failure paths for an engineer to judge. The engineer stays in charge; the aim is support, not replacement.
Theme lead: Dr Zhibao Mian
Assistant Professor · Lecturer in Computer Science
Leads the AI for Safety theme, working on AI-assisted analysis, prediction and diagnostics for engineered systems.
Also contributing
AI-assisted safety analysis
Using learned and generative methods to propose hazards, failure modes and chains of cause that a manual analysis would likely miss, then putting them to engineers for review rather than accepting them automatically.
Prediction
Anticipating wear and failure from operating and condition-monitoring data, so that people act before a hazard happens rather than after an alarm.
Diagnostics
Working back from an observed symptom to the component and mechanism responsible, in systems where many causes look alike and the diagnosis has to be defensible.
Decision support
Giving operators and safety engineers a sound basis for a judgement — what’s happening, how confident the system is, and what the options are — instead of an unexplained recommendation.

Open questions
There’s an obvious circularity here, and we take it seriously:
• If AI produces part of a safety analysis, what assurance does that AI itself need — and does theme 01 have to be solved first?
• How should engineers review a generated fault tree so that over-trust in automation doesn’t quietly replace their judgement?
• What evidence would a certification body accept for an analysis that a person didn’t write?
Safety analysis not keeping up? We work with partners whose systems have outgrown manual hazard analysis and maintenance planning.
Related themes:
04 · Model-Based Safety & Dependability
06 · Safety and Assurance Tools