Building Trustworthy AI for Pipeline Operations: Insights from Klarian's Lead Data Scientist

It’s easy to view the implementation of these systems as a purely technical conundrum, but to develop tools that actually create an impact, you need a holistic approach. The only way to achieve this is by building a development process that keeps the user in mind from beginning to end. Working alongside AI is already a reality for many in the pipeline sector, and Anna maintains that this relationship will be the most productive when AI tools are built to fit into existing workflows.

Here are Anna’s thoughts on how AI can have a positive impact on the pipeline sector.

What is the aim of your work?

My work is focused on developing the AI that enables Juno’s decision support capability. Juno helps pipeline operators gain better insights into their operations, combining physics-based modelling and AI analytics to unlock better performance across a number of key metrics.

I like to imagine three horizons when it comes to decision-making. The first is the reactive horizon. This kind of decision-making needs to be done quickly and typically exists to address immediate problems. It is a part of pipeline operations, and teams need to act effectively in this space. However, spotting problems and opportunities earlier, along with making more considered decisions, tends to lead to better results.

The next horizon is tactical decision-making, which focuses on what operators can do to optimise their systems. It encompasses predictive maintenance, spotting problems before they materialise and preventing them. This isn’t necessarily about fixing things that have broken but improving what’s already there. These decisions concern medium-term improvements to operations and course corrections to avoid issues.

The final horizon is the strategic horizon, which allows operators to make more impactful decisions around the underlying system. For example, Juno was used by the British Pipeline Agency, an operator jointly owned by Shell and BP, to investigate a hydraulic throttling issue. Juno’s insights were used to build a business case to invest in a new variable speed drive pump. To make impactful, strategic decisions, operators need access to the insights that Juno provides.

Success in the pipeline sector depends on making impactful decisions across all three of these horizons. However, it is often easier to make better-informed decisions in the tactical and strategic spaces. That is the focus of my work on Juno: to give operators better foresight and more information on their pipeline performance so they can make more tactical and strategic decisions for better overall results.

How does Klarian approach AI?

One of the things that we’ve implemented – and that I’ve drawn upon from my background in academia – is a scientific approach. I’m a data scientist. The distinction I draw from that title versus similar titles is that I use the scientific method to solve problems. This means that all of our work is hypothesis-driven and that peer review is built into the process.

Another important part of our approach is building trust. Developing AI brings a lot of technical challenges, but building trust with a user is truly make or break for any tool. You have to make sure that what you’re putting out into the world is reliable and solid.

When building systems, it’s one thing to make them accurate and reliable; it’s another challenge to build the interface in a way that is explainable. For example, when displaying data about pipeline operations and the trends that emerge, it’s important to provide users with the chance to drill down into that data to understand what’s happening. The ideal scenario is where reliability and trust are so high that the need to use this functionality reduces over time. Of course, we will maintain this functionality even in this case as explainability is a key principle of our approach.

What impact can AI have on pipeline operations?

For most people, AI means tools such as ChatGPT and other large language models. These types of AI are built using massive datasets. There’s no shortage of text on the internet for them to ingest. For the types of AI that Klarian uses, we don’t always have the privilege of near-endless datasets. Pipeline operations are complex, and the conditions change regularly. This can create data sparsity – which is a real challenge.

There is no ‘one-size-fits-all’ approach when it comes to selecting a type of model. I approach each problem in a systematic way, beginning by testing simple solutions, recognising their limitations, and then adopting a new solution, building complexity gradually. This ensures we can be accurate and find the flaws in our models early so we can fix them. Overcoming sparse data requires us to be creative with our models. We can do things like aggregation to improve our predictions by looking at the whole system. The principle remains the same: start simple and add complexity over time as we gain confidence in our approach.

It’s easy to view the challenge of sparse data in purely technical terms, but it goes further than that. We have to consider the people who will use our tools, and by bringing them into conversations as early as possible, we can build better systems that solve the correct problems.

One use case for Juno is in the area of uncommanded pump stops. These stops cause downtime, and we want to provide quicker, better root cause analysis on the reactive time horizon while offering aggregated insights to guide tactical and strategic decisions. We have found a few characteristic system behaviors, or ‘pump stop personas,’ that lead to these stops. This allows us to present quantitative analysis of the stops, giving operators more insight into how and when stops occur. Over time, Juno will recognise the patterns associated with a stop and flag potential pump stops, moving analysis from a reactive to a proactive space.

What factor is most important for developing AI in the pipeline sector?

Having the right team.

At Klarian, we have a team built around the values of trust and reliability. Everyone is empowered to speak up if they spot something that’s not right. We work in a customer-centric way, and that means that we first make sure we’re asking the right questions and then ensure we arrive at the correct answers.

Part of the scientific method is peer review, and we’ve built this into each step of our process so that our customers can rely on Juno’s predictions and recommendations. Training our people to work in this way instils the values of trust, reliability , and being dependable. If those values exist in our people, they are transferred into our technology, and our customers feel the benefit.

Conclusion

Anna has been, and continues to be, heavily involved in the development of Juno. The software works as she describes, using trusted, reliable means to solve problems. AI plays a large role in what Juno does, but our approach is always explainable, and Juno is developed to use the right tools for the job. That means processes that require AI will use AI, but where Juno can use deterministic means, like physics-based modelling, it will do so.

Mixing these methods, using a hybrid approach, is what delivers a 20% uplift in efficiency for pipeline networks. Anna’s work on Juno helps pipeline operators gain deeper insights and see further into the future to better strategise around their pipeline’s performance. Anna is a key part of the team driving Klarian forward and helping to deliver value to each of our customers.