Nuclear is ready for Artificial Intelligence
Nuclear has not yet taken the big step into using Artificial Intelligence (AI) and it’s easy to see why. One is an industry that moves at a cautious pace and is focused on safety, while the other is shaped around building tools in agile sprints by failing fast.
Now, however, is the perfect time for the nuclear industry to start embracing AI and using it to revolutionise existing processes, adopting data science to help engineers, not replace them. By applying the right blend of sector experience and data science expertise, these decision support tools can help improve the safety culture and achieve significant operational and cost benefits.
In short, more data science leaves more time for real engineering. So where can AI really help?
Data rich, information poor
Many organisations’ data is not necessarily stored in a machine-readable format, some is not even in a database. Even today, large volumes of data are stored as text data in documents which are very difficult to analyse quickly and comprehensively. To compound the problem, these documents are not necessarily written in a way that is easy to digest without having specific hands-on experience. Focusing on nuclear safety, this makes it harder to identify and address gaps in our existing safety cases.
By developing and deploying AI based solutions to automate and enhance these manual yet specialised processes, we can start to achieve immediate tangible operational and safety benefits. For instance, we can reduce the time taken for desk studies from months to minutes, avoiding the need to limit and prioritise the safety documents that are in scope. We can now develop solutions that can support both the development and verification stages of safety cases.
Taking this one step further, these solutions can help us tackle tasks and answer questions we’ve never been able to do before. For example, if you want to understand the commercial and safety implications of removing a system from the rest of the plant, the first step is to map all existing dependencies. However, this task could take an engineer up to an entire year to review all relevant documentation, so alternatives are generally preferred (e.g. conducting a workshop with relevant experts). These alternatives, though, can lead to gaps which are subsequently raised at a verification stage and can cause a substantial amount of work and rework.
Example of how we can use AI to draw connections between key systems based on the plant’s underlying documentation. If one is interested in a specific system, we can use this network to highlight all its key dependencies (e.g. can we shut off and stop maintaining this system during defuelling or does it support other key systems?)
By streamlining our ability to answer complex questions, we ultimately become better engineers. It is our critical thinking and engineering judgement which forms the basis of our value, not our ability to retain knowledge that spans across 1,000+ documents.
The table (below) shows some of the examples where an AI solution could provide significant operational and cost benefits, whilst also improving safety. AI based solutions have the advantage of always remaining up to date and can easily be scaled up. The first four use cases focus on how AI can be applied to safety documents, whilst the bottom four provide applications that stretch to wider processes.
In time, we can easily foresee an engineer getting near-instantaneous tailored responses or reports by simply using a free text box to enter a range of specialised questions. For instance:
What are the safety claims associated with the Charge Machine?
Which systems would be impacted if we stopped using the buffer stores?
What documents describe the Flask Hall Crane’s operations?
Who are the top contributors (authors and verifiers) to the Irradiated Fuel Dismantling Facility’s Safety Case?
The figure above shows how we can use AI to find which specific references underpin each system
The nuclear industry has always actively shared lessons learned in its day-to-day operations. However, this still relies on its experts using traditional means of communication.
A step forward is to leverage AI solutions to process and analyse our documents to provide new ways of sharing knowledge that is faster and data driven. This provides a new and innovative way of connecting a nuclear fleet together. For instance, we can help our engineers compare historical and future plant modifications, or identify the impact different maintenance regimes may have on the performance of each plant.
So how does AI work?
When looking at text data, the key terms to know are Machine Learning (ML), and Natural Language Processing (NLP). These are important subsets of AI. ML models iteratively learn from data and can find hidden insights without being told where to look, while NLP is used to analyse and process unstructured natural language data (text and audio).
Machine learning’s long serving nemesis has been what is known as ‘knowledge engineering’. The idea is that ‘knowledge’ cannot be learned automatically and must be programmed into the computer by human experts. However, the statistical laws that underpin ML and NLP algorithms can help connect the dots much better than we would be able to.
For instance, we may think that our experience means we understand nuclear documents better, but we are in fact training our own ‘biological machine-learning model’ to look for key terms or recognise specific descriptions. Think about how you recognise knives in a drawer —you don’t need to have seen every type of knife in the world to correctly identify one.
There are several established NLP techniques and pre-trained models that can help automate manual activities with minimal computational cost. These techniques can be used to automatically ingest, process and extract information (e.g. key systems or hazards) from thousands of documents. They can do so much faster than even an experienced engineer could and therefore allow risks to be identified and managed rapidly.
Humans look at words and machines look at numbers. NLP models translate words, sentences and documents into vectors to create structure out of the inherently unstructured natural language text data. This means we can develop solutions that do not simply look at relative term frequency to determine its importance. Instead, they can consider other factors, such as the context of surrounding words on the meaning of individual words. We can then use these distributed numerical representations of word features to mathematically detect similarities between words, sentences and documents.
An example of how semantic similarity can be used to group sentences together; rather than searching for key words, these techniques analyse the underlying context.
Theoretically, given enough data usage and contexts, algorithms can make highly accurate suggestions about a word’s meaning based on past appearances and the context of surrounding words. These suggestions can be used to establish a word’s association with other words.
Example of how you can use text similarity to find dependencies between systems and help visualise and understand how they suport each other
Can it work in nuclear?
The nuclear industry is beyond ready for data science innovation. There are three key areas that are important to examine: data, infrastructure and culture.
Both data and infrastructure’s current forms are such that adopting even the smallest of AI solutions can lead to a huge operational benefit. If we couple that with a culture that has safety as its overriding priority and an emerging workforce who are entering the sector with high computational proficiencies, the nuclear sector is very well positioned to become a leader in this field.
When looking at data; the garbage in, garbage out headline used to be the go-to worry. Today the headline issue is bias. Whilst these are important considerations, they are not showstoppers, and so the idea that nuclear data is too specialised to work with AI is incorrect. Analysing nuclear data will present its specificities, but it is no more challenging than analysing contractual data to identify potential cost savings or predicting customer churn using customer survey responses.
There is, however, a problem with that fact that the technological infrastructure the industry uses has remained the same for decades. The problem is not with the tools themselves, but with its users pushing them far beyond what they were built for. Currently, excel spreadsheets are being used as databases, dashboards or even to run complex models.
For some nuclear organisations, the coronavirus pandemic was the catalyst that enabled mass roll-out of communication platforms, enabling cross-organisational conference calls. This shows that if appropriate, our sector can implement new technology to help day-to-day operations. However, waiting for the next low probability event/disaster to ignite the next technological breakthrough might not be the best strategy.
An example of how ML and NLP can be used to move away from Excel and develop new ways of to extract and visualise text data. There are data science techniques that help us classify a document into discrete topics. Each topic consists of key terms and is represented by a bubble.
The culture challenge lies in the way we put safety first and so have stopped exploring new solutions and have become too reliant on our existing technology. Instead of looking to innovation, we have favoured specialised training courses and company specific certifications to provide trainees with tips and shortcuts that unfortunately cannot scale up. Instead, we should learn from Central Government, which is already building self-service applications to encourage its users that they no longer need to download data and throw it into a spreadsheet in order to derive insights. Adopting innovative solutions will fundamentally augment our ways of working and radically improve our accessibility to data and knowledge transfer between engineers.
These solutions do not hinder safety, they improve it.
How would we begin?
Nuclear organisations have many ideas for potential data innovation but often find it difficult to navigate through the evolving technology landscape. In addition, mobilising the right capabilities and resource to explore these ideas can take significant time and effort.
Our industry involves complex systems which have undergone substantial historical modifications and are underpinned by its comprehensive safety documentation. But as the volume of processes and documents increased, we increased our workforce but made no changes to our infrastructure or our ways of working. This means we are still doing tasks that machines are now much better equipped to handle.
To ensure the use of AI has an immediate impact, we need bespoke solutions which are designed to specifically solve our most pressing problems. Whether a solution combines a wide range of techniques or consists of a single proprietary product is irrelevant, what matters is that its core design is tailored to a specific need.
An important first step is to set your data strategy to ensure both vision and level of ambition are clear and shared across the organisation. Just because an organisation hires data scientists, engineers and front-end developers does not mean it can suddenly develop and deploy successful AI based solutions. Without a clear vision, an organisation could end up saturating its digital eco-system with countless Proof of Concepts.
Secondly, it is critical to conduct discrete and targeted discovery phases ahead of starting an innovative transformation project. Simply because executives and managers sit through an AI course, it does not mean they are able to recognise where the best opportunities for innovation are. This activity requires the combination of sector experience and data science expertise to be able to jointly identify the innovative opportunities required to solve the most pressing business challenges.
As part of our culture we are encouraged to challenge our thinking. However, it’s time to embrace our enthusiasm for innovation and focus on the good ideas — not just the good challenges.
We often jump to hypothetical reasons why an AI solution might not work. In reality, we are simply listing its potential limitations, just as a person has their own limitations in the amount of activities they can undertake. Our ability to apply our critical thinking and engineering judgement to improve nuclear safety can only get better, provided we start using the right tools.
The way forward for the nuclear industry is to start adopting data science to revolutionise our existing processes, and help our engineers increase safety whilst reducing cost. It’s time to think big, start small and scale fast!