New software tool shows clear advantage in water purity prediction
Published: 7 October 2025
A powerful new software tool which can accurately predict the performance of biofilters used by the water industry could reduce the challenge of maintaining the purity of tap water.
A powerful new software tool which can accurately predict the performance of biofilters used by the water industry could reduce the challenge of maintaining the purity of tap water.
Researchers from the University of Glasgow’s James Watt School of Engineering developed the tool, called the Environmental Buckingham Pi Neural Network, or EnviroPiNet.
It uses machine learning techniques paired with sophisticated physical modelling to predict the ability of biofilters to remove organic carbon compounds from water with up to 90% accuracy. The tool is now available online for free use.
Biofilters are used in water treatment facilities to remove contaminants from wastewater before it is returned to domestic taps as pure drinking water. Organic carbon compounds are a common pollutant, finding their way into water through routes including human waste, agriculture and industry.
Biofilters are an environmentally-friendly method of removing carbon and other contaminants from water. They are coated in layers of bacteria which capture and consume carbon from wastewater as it passes through. Their performance degrades over time as waste material builds up, meaning biofilters need to be monitored closely to ensure water quality doesn’t degrade along with them.
Being able to accurately model the performance of biofilters in real-world applications could help the water industry make better-informed decisions about how they are used and when they are replaced. However, the development of software tools to help with these tasks has been held back by a lack of high-quality data about the complex physical environment of wastewater treatment facilities.
In a new paper published in the journal Scientific Reports, the team show how EnviroPiNet combines a mathematical theorem called Buckingham Pi with machine learning to make accurate predictions about biofilter performance despite the sparse data available.
The team compiled a dataset of biofilter performance drawn from previous research literature and their own lab tests, which they used to train the machine learning algorithm. They used 80% of the data available to teach the model, and, once trained, validated its performance by testing its ability to model the data in the remaining 20% of the dataset.
EnviroPiNet was able to accurately predict biofilter performance 90% of the time, a clear improvement over other models the team tested using the same dataset. The second-best performance using PCA-based methods achieved 50% accuracy, while autoencoder techniques achieved just 20%.
Uzma, of the University of Glasgow’s James Watt School of Engineering, is the paper’s corresponding author. She said: “Environmental biotechnologies present unique challenges, as they involve complex microbial systems and often depend on data collected under limited or controlled conditions. This lack of diverse, high-variability data can make predictive models less generalisable and reduce their accuracy in real-world scenarios. Even with modern high-resolution monitoring methods, challenges with data quality, integration, and representativeness persist, highlighting the need for improved data strategies to enhance model performance.
“EnviroPiNet is the first tool of its kind to demonstrate highly-accurate predictions from the data available, which is a really encouraging result. We are currently working with partners in the water industry to find ways to test EnviroPiNet’s performance in real-world conditions. Our hope is that it will be a valuable tool to help fine-tune the performance of biofilters to maximise their performance without costly and complex physical tests, reducing costs and increasing efficiency without compromising water quality.”
The team are already exploring how EnviroPiNet might be adapted to other uses, including in healthcare settings.
Dr Fabien Cholet, Domenic Quinn, Professor Cindy J Smith, Professor Siming You, and Professor William T Sloan of the James Watt School of Engineering are the paper’s co-authors. The team’s paper, titled ‘A machine learning model guided by physical principles for biofilter performance prediction’, is published in Scientific Reports.
First published: 7 October 2025