Defining a path to safer and more sustainable cooling
Savantic
When CERN needed to explore new ways of monitoring and controlling Legionella growth in its evaporative cooling towers, Swedish AI company Savantic AB conducted a pre-study to assess feasibility and outline a model-based approach combining physics and machine learning.
What we develop for CERN’s cooling towers can also benefit hospitals, industry and real estate, opening the door to a much larger market.
The collaboration began at AIMday Big Science Technology 2025, an event dedicated to green tech solutions for large-scale research facilities. The cooling tower problem was raised there, and Savantic engaged directly with CERN stakeholders to find a solution.
From reactive testing to predictive control
Cooling towers are essential to CERN’s infrastructure, dissipating the vast amounts of heat generated by particle accelerators, detectors, cryogenic systems and data centres. However, these environments also create favourable conditions for Legionella pneumophila, posing both health risks and regulatory challenges.
The screening method that is mandated by French regulation relies on culture-based laboratory analysis, which can take up to two weeks to deliver results. Operators therefore apply chemicals such as biocides and corrosion inhibitors on a precautionary basis rather than according to actual need.
Savantic’s study identified a clear opportunity: shifting from calendar-based maintenance to predictive, risk-based control without compromising regulatory compliance.
Combining sensing, modelling and AI
Savantic conducted a comprehensive, vendor-neutral assessment of rapid detection technologies, including qPCR, antibody-based methods and nanopore sequencing. In parallel, the company reviewed the state of the art in modelling bacterial growth.
Following this, Savantic proposed a modelling approach built on interpretability rather than a black box:
A bio-physical growth model, consisting of a logistic growth equation, driven wherever possible by process parameters the towers already measure.
A neural-network alternative, benchmarked against time-series models, should the physical interpretation prove insufficient.
An ancillary corrosion and scaling model, since pipe condition determines the habitat available to the bacteria, and since corrosion inhibitors are the largest driver of environmental impact.
The approach also includes a structured plan for data collection, calibration against mandatory culture testing, and a 15-month implementation roadmap.
A key insight is that rapid testing technologies are primarily needed during the initial calibration phase. Once validated, the model can run on continuous process data, complemented by the mandated regulatory sampling.
Reducing cost, risk and environmental impact
The proposed solutionificant benefits for CERN and similar facilities:
- Environmental impact: reduced use of biocides and corrosion inhibitors by 10%, potentially more.
- Cost efficiency: estimated chemical cost savings of 10–50%.
- Work environment: reduced handling and exposure to hazardous chemicals.
- Operational reliability: early warnings lower the risk of costly downtime.
- Scalability: transferable to other cooling systems and suitable for federated learning across sites.