New AI Tool Predicts Chemical Removal in Wastewater Treatment
Sabine Barrett, SETAC
A new study published in Integrated Environmental Assessment and Management introduces an artificial intelligence tool that could significantly change how scientists evaluate the environmental fate of chemicals. Developed by researchers at Environment and Climate Change Canada, the Canadian A.I. Removal Rate Estimator (CAIRRE) offers a novel way to predict how effectively chemicals are removed during wastewater treatment—an essential step in assessing environmental risk.
The study, led by Thomas D. Burns and colleagues, responds to a growing challenge in environmental science: the rapid increase in new chemical substances entering the market. While innovation in chemistry has brought advances in medicine, materials and technology, it has also outpaced the ability of regulators and scientists to fully understand the environmental impacts of these substances. For wastewater treatment plants – where many pollutants are either removed or released into the environment – predicting how chemicals behave remains a persistent gap in risk assessment.
Traditionally, scientists have relied on mechanistic or “fate-based” models to estimate removal efficiency in wastewater treatment plants. These models simulate how chemicals partition, degrade or persist, but they depend heavily on detailed input data, such as physicochemical properties and biodegradation rates. In many cases, especially for newly developed substances, those data are incomplete or unavailable. As a result, predictions can vary widely, introducing uncertainty into regulatory decisions.
CAIRRE takes a different approach. Instead of requiring extensive input parameters, the model uses artificial intelligence to predict removal efficiency directly from a chemical’s structure. By training on a large dataset of measured removal rates from wastewater treatment plants in Canada, California and other regions, the model learns patterns that link molecular features to real-world outcomes. This allows it to estimate how a substance will behave in a typical secondary wastewater treatment system, even when little experimental data exist.
The research team compiled more than 18,000 observations from 182 wastewater treatment plants, ultimately refining the dataset to 161 chemicals with sufficient data for model training. Using a machine learning method known as gradient-boosted decision trees, CAIRRE achieved strong predictive performance. Validation results showed that the model could explain a substantial portion of the variability in removal efficiencies and that most predictions fell within a reasonable range of observed values.
The model also outperformed widely used existing tools when tested under comparable conditions. These traditional models often struggle when input data are uncertain or incomplete, whereas CAIRRE avoids that limitation by relying solely on chemical structure. In practical terms, this means risk assessors may be able to generate more reliable predictions earlier in the evaluation process, particularly for data-poor substances.
As an emerging tool, CAIRRE is not intended to replace existing models, the authors note. Its predictions reflect generalized conditions and do not account for site-specific factors, such as temperature, pH or plant design, that can influence removal efficiency in practice. The model is also trained primarily on data from North America, which may introduce regional bias. As such, it is best used as a complementary approach, particularly in early-stage assessments or when data are scarce.
At the same time, environmental science is increasingly shifting toward data-driven approaches as the number of new chemicals continues to grow. By leveraging large datasets and machine learning, tools like CAIRRE can help close data gaps and improve the speed and consistency of chemical risk assessments. As these models evolve, the authors suggest AI could play an increasingly central role in protecting environmental and human health.
Read the full article at https://doi.org/10.1093/inteam/vjaf163.
Contact: [email protected]