How can artificial intelligence help map PFAS high-risk areas?

Is AI (artificial intelligence) more than large language models, enabling the generation of pictures and the preparation of emails?  

COWI's groundwater professionals have challenged the use of AI and returned to a more fundamental approach which uses machine learning to get a deeper understanding of complex structures in data. Our experts have been investigating whether machine learning could help reveal structures in groundwater data that would work as a holy grail for understanding and handling groundwater contamination with complex and potentially very harmful contaminants, such as PFAS. 

PFAS (per- and polyfluoroalkyl substances) consist of a large group of synthetic fluorinated substances that have been used in different industries since the early 1950s. These substances are stable, difficult to break down and highly water-soluble. In everyday language, these contaminants have therefore become known as 'forever chemicals'.  

Contamination data requires a dynamic approach 

Current estimates indicate that up to 15,000 sites in Denmark might be contaminated by PFAS due to past activities. This presents a dilemma for the five regions of Denmark (the authorities responsible for contaminated sites in Denmark) regarding how to best prioritise resources for investigation and remediation activities. Similarly, these ‘forever chemicals’ challenge Danish water supplies, whose primary task is to produce and deliver clean drinking water to Danish consumers. 

By applying machine learning, COWI has investigated the potential of integrating data from, e.g., the hydrogeological realm and groundwater measurements of PFAS into a complex network where internal correlations between data illuminate important structures or dependencies. Through this exercise, our experts have been able to answer a long list of questions, such as which of 11 defining parameters – e.g., the land use, infiltration or the thickness of geological layers – is more important to the presence of PFAS in a groundwater aquifer. On the basis of the data, a model is trained to predict the vulnerability of different groundwater aquifers toward PFAS contamination in areas where PFAS measurements have not been done, but where valuable hydrogeological information is available. The results are visualised on a map of 13 different groundwater aquifers in Denmark, and different stakeholders, both public and private, are offered free access to the maps through publicly available websites. 

Being dynamic, contamination data requires a dynamic approach. As more and more water samples are analysed for PFAS, a static model would quickly become outdated. The project has, consequently, pursued a dynamic approach which lets the model retrain itself with minimal intervention. This means that the model automatically draws on new available data directly from external databases. Hence, the tool for prioritising the ‘forever chemicals’ will remain evergreen. 

Fake news or reliable helpers?  

It is, of course, crucial to be sceptical of new technologies, but the strength of machine learning is that it is possible to measure how credible these qualified guesses are. By taking one fourth of the data for validation purposes, it is possible to compare the guesses with reality. Thereby, the experts can compare the guess with the measured value at a given point, thus providing a measure of how accurately the model guesses. For the PFAS model, 83% of the predictions are shown to be correct, which is quite good within machine learning. The model also excels at identifying locations where PFAS has been detected, therefore making both accurate and precise predictions, which results in credible outcomes. 

Since the model is based on publicly available data, the results are being integrated into publicly available websites to achieve broad access and ease of use for various stakeholders. With this approach, the tool can be used by regions, suppliers, municipalities and private entities, with the hope that the model can serve as a valuable tool for assessing PFAS vulnerability and support decision-making processes in environmental protection and groundwater extraction. 

This project is an example of the innovative use of publicly available data to solve complex environmental problems. With the continued development and broader application of machine learning and artificial intelligence, COWI sees great opportunities to employ these technologies to better protect our groundwater and the environment as a whole.