Scientific machine learning is key to re-engineering maintenance and reducing process emissions across water operations, says Tom Ray, director of digital products & services at Binnies UK.
As ageing infrastructure meets tightening environmental standards, the water sector needs smarter, more predictive ways to operate – improving performance while cutting emissions at source.
The breakthrough will come from looking beyond the sector to how advanced AI, such as scientific machine learning, is already transforming performance and reliability in industries such as aerospace, energy and motorsport.
A poll at the 2025 industry Predictive Asset Health Summit – held at the Atlassian Williams Racing (F1) HQ – revealed that more than half of all maintenance activity in the water sector remains reactive, triggered by a fault rather than prevented by foresight.
As unplanned failures and emissions compliance pressures rise under AMP8, this approach will become increasingly unsustainable. Both the Cunliffe Review and Ofwat’s Asset Health Roadmap have called for a decisive shift towards proactive, data-driven asset management.
Studies – including one from the US Department of Energy – show that predictive maintenance can deliver major benefits, generating up to a tenfold return on investment through reduced downtime, optimised performance and lower operational costs.
Yet for water companies, achieving truly predictive maintenance across complex treatment and pumping systems remains difficult. Many assets are sparsely instrumented, with operational data that is incomplete, unavailable or noisy. As a result, many have found that traditional AI techniques struggle to produce reliable or transferable insights.
Traditional machine learning performs well when abundant, high-quality historical data is available for training. In practice, that often requires large-scale and costly sensor deployment programmes before models can deliver useful results – a significant financial and operational barrier for many utilities.
Even then, traditional machine learning identifies correlations rather than causes. It recognises statistical patterns but not the underlying physics, chemistry or biology that drive those patterns, limiting its ability to contextualise or adapt when conditions change.
Scientific machine learning takes a fundamentally different approach. It combines these scientific principles directly with machine learning to create models that inherently ‘understand’ the physical and biological processes at play.
Designed for complex engineered systems – from aerospace and aviation to energy and elite motorsport – scientific machine learning was built specifically for environments where traditional machine learning struggles to scale due to process and data complexity.
With scientific machine learning, the first question is not “what data do you have?” but “what engineering information exists?”
A pump curve, operations and maintenance manual or design information can often provide around 80% of the baseline model fidelity before any sensor data is introduced. The model is then calibrated using available telemetry – such as flow, current or temperature – and continuously refined through machine learning and neural networks.
Where data is missing, scientific equations can infer it, effectively creating virtual sensors that fill information gaps and reduce – or even remove – the need for additional hardware. The result is a model capable of predicting asset health and process performance accurately, even in data-limited environments.
This approach is already showing strong results.
At Southern Water’s Matts Hill Water Supply Works, a digital twin was created for the site’s borehole pumps.
Using only four sensor inputs – voltage, current, flow and well level – combined with manufacturer pump curves and site configuration details, Binnies, in partnership with Williams Grand Prix Technologies and JuliaHub, developed a virtual model powered by scientific machine learning that predicts pump faults with over 90% accuracy.
The same model also identified energy-efficiency losses caused by changes to control settings over time, enabling operators to restore optimal performance and reduce power consumption – insights that would otherwise have required extensive on-site monitoring.
On the environmental-performance side, a particularly promising application is in reducing nitrous oxide emissions from activated sludge plants – typically the single largest source of process emissions across wastewater treatment works.
By building a physics-based digital twin of the plant, scientific machine learning can simulate the complex biological and chemical interactions within aeration tanks, accounting for geometry, aerator configuration, loading rates and process equations.
The model learns how these variables interact to influence nitrous oxide formation, enabling operators to identify the optimal balance between aeration efficiency, treatment performance and emissions reduction.
This provides predictive guidance on both emissions and energy use in near-real-time, supporting progress towards Ofwat’s 2030 price control deliverables for net zero (PCDWW34) and compliance with the Industrial Emissions Directive.
Scientific machine learning delivers the greatest value on complex, dynamic systems that generate measurable operational signals – such as critical treatment pumps, pumping stations, rising mains and sludge-treatment centres.
It is less suited to static assets with little or no dynamic behaviour, such as weir gates or balance tanks, which are unlikely to deliver the same level of return on investment.
While still new to the water industry, scientific machine learning has already proven itself in aerospace, energy and elite motorsport. It offers a breakthrough in how we understand, predict and optimise both maintenance and emissions performance across water operations.
Tom Ray is the director of digital products & services at Binnies UK, where he leads the introduction of digital twins and scientific machine learning into the water sector through a partnership with Williams Grand Prix Technologies and JuliaHub.
