Large Language Models for Civil Nuclear Site Performance Improvement

Sellafield is one of the world’s largest and most complex nuclear facilities, staffed by over 11,000 personnel and comprising over 500 facilities dating from the 1950s to the present day. Site activity ranges from interim storage of spent nuclear fuel to clean-up of legacy facilities, alongside all the services, business processes and security considerations that come with operation of a large, sensitive industrial facility.

In support of Nuclear Site Licence Condition 7 - ‘Incidents on Site’, Sellafield raise and manage Condition Reports (CRs) via the Performance Improvement (PI) process. CRs are free text data, relating to any ‘off normal condition’ – this can range from conventional safety hazards or injuries, to plant performance and equipment reliability issues. CRs can be raised by anyone within Sellafield or involved in Sellafield activities.

These CRs contain valuable insights which can be used to drive performance improvement across site operations and each CR is therefore reviewed, categorised and investigated/actioned appropriately. 'Trend Codes' are applied to each CR, grouping them into themes to help derive insights. However, trend coding of CRs takes a lot of time.

Ada Mode therefore investigated the feasibility of using an LLM to automate CR trend coding and provide recommended codes to the performance improvement team. The aim was to demonstrate that an LLM can achieve human-level performance in this area and can ultimately be integrated into day-to-day operations to dramatically speed up CR processing time and reduce any coding back-log. Data security was a major consideration here, preventing the use of established online LLMs like ChatGPT, and preventing the use of cloud computing services.

Ada Mode built specialised hardware to enable secure on-premises LLM training and tested a range of LLMs to appropriately balance accuracy with computational requirements. It is critical in on-premises deployments to ensure that care is taken to keep hardware needs manageable. Despite this constraint, we were able to deliver a model which matches human-level performance and is able to trend code a CR in << 1 second.

In parallel with LLM development, Ada Mode built an interactive dashboard to visualise CR data to maximise ability to recognise trends and drive performance improvement.

Having demonstrated that LLMs can reliably process CRs, work is now moving into design and implementation an upgraded overall CR trend coding process.

LLMs are extremely powerful at extracting meaningful insights from free text data while accounting for context and domain-specific language. However, data security is a particular challenge which currently limits potential LLM applications for many businesses. This is particularly the case in civil nuclear where data is often subject to formal security classifications. As such, getting value from LLMs within nuclear will often require use of secure on-premises or self-managed cloud-based applications.

You can read more about our work with and thoughts on LLMs here.