Machine Learning on Track: enabling predictive maintenance and smarter decision-making

TL;DR:

At Rail Live 2025, Network Rail showed how Machine Learning is reshaping rail operations - from predictive maintenance to smarter decision-making. At Atmo, we use Machine Learning to reduce operational waste, ensure safety, and unlock insights from forgotten data. Intrigued? Read on, it will only take you 3 minutes.

-

Insights from Rail Live 2025: Machine Learning in Network Rail

Inspired by the insightful discussion between Martin Mason (Senior Programme Manager, Network Rail) and Gareth Denis (Strategy Lead, Test Tracks, Network Rail) at the Rail Live 2025 conference last week, we felt it is imperative to highlight the power of learning using AI. But what does that really mean? Let’s break it down.

In the spirit of Rail Live introspection, let’s revisit an example Martin and Gareth shared during their talk. By harnessing the power of machine learning, Martin’s team has developed a model that identifies patterns and transforms failures/learnings into opportunities. These opportunities are captured through their predictive maintenance tool, enabling smarter decision-making driven by better data.

From solving one problem to addressing multiple

Machine learning has shifted Network Rail’s approach from solving one problem at a time to addressing multiple issues simultaneously. Through simulation and predictive learning, the team can pinpoint root causes, distinguish between true and false positives, and gain a deeper understanding of persistent issues. This, in turn, drives improvements in performance, extends asset lifespan, and enhances staff safety.

“re-analyse the state of play in real time and come up with a new decision (asking the much-needed question each time), is right now exactly the right moment to do something?”

As our lead data scientist, Fynn O’Connor, puts it, machine learning allows us to “re-analyse the state of play in real time and come up with a new decision (asking the much-needed question each time), is right now exactly the right moment to do something?” He adds that machine learning enables us “to pre-empt something you might not have anticipated in the pre-AI era.”

Quality Data is key

We’re bombarded with data — some of it makes sense, much of it doesn’t. 

All of this is powered by data. The model is trained on the data we provide, and by embracing a growth mindset that Martin advocates for, we free ourselves from the fear of the unknown. Too often, we build workflows based solely on our current knowledge and processes. We’re bombarded with data — some of it makes sense, much of it doesn’t. Machine learning helps us, first and foremost, to make sense of that data, but more importantly, it empowers us to embrace failure as a learning tool.

Through simulations, we can quickly learn from past lessons learned and use those lessons to build a safer, cleaner, and more efficient industry, while gaining the ability to pre-empt issues before they arise.

Creating a Resilient System

That’s not to say machine learning will solve everything. Our models are only as powerful as the data and policies that guide them. The human mind is essential, to feed in high-quality data, and to design integrated and robust systems that ensure resilience. Fynn often reminds us at Atmo to be vigilant about erroneous or missing data. He also emphasizes the importance of having a backup plan (in the event of a power failure), the system must be able to withstand disruption and not rely solely on a constant data feed to function. 

Machine Learning at Atmo

So, how do we harness the power of machine learning at Atmo?

We use a subset of machine learning, specifically, Large Language Models (LLMs), to extract valuable insights from old and often forgotten reports. Think about those expensive reports that are in your desk drawers that were often a huge investment to produce. These reports contain historical data, which we use to train our models. Leveraging deep learning, our models then optimise decision-making by identifying the most effective course of action based on that data.

This process is refined over time, as the model continues to learn and improve with each new data input.

Real-world Machine Learning in Rail

Okay, this all sounds great — but how does it apply to the real world?

From our work at Atmo, we’ve identified a major challenge in the industry that up to 70% of resources used for dust suppression at rail aggregate material sites are wasted due to manual or timer-based systems. Our machine learning model addresses this inefficiency by enabling smarter, data-driven control of dust suppression systems, which in turn, reduces costs and conserves valuable resources like water and energy for our Atmo clients. This all feeds back to our relentless drive to develop technologies that contribute to a safer, cleaner, and more efficient industry - whether in rail, construction or nuclear.

Curious to see what this could look like for you and your team? Our data scientists are always happy to chat and show you how our tech can help you optimise your safety & environmental operations. 

Say “Hi” at hello@atmotech.co.uk