AI and IoT promise a closed-loop future for tracking persistent pollutants

Persistent organic pollutants, from polychlorinated biphenyls and organochlorine pesticides to brominated flame retardants and per- and polyfluoroalkyl substances, are among the most stubborn legacies of industrial development. Because of their chemical stability, resistance to degradation, and capacity to bioaccumulate, they linger in the environment for decades, cycling among air, water, soil, sediments, plants, animals, and humans. Long-range atmospheric transport and biomagnification carry them far beyond their original sources and up the food chain. A growing body of epidemiological and toxicological evidence links chronic exposure to endocrine disruption, carcinogenesis, reproductive impairment, neurodevelopmental disorders, immunotoxicity, metabolic disease, and biodiversity loss. Although the Stockholm Convention has curbed production of many legacy compounds, secondary emissions from contaminated soils and sediments, illegal or exempted uses, newly recognized contaminants, and climate-driven remobilization from glaciers, permafrost, and aquatic sediments keep reintroducing these chemicals into the environment.

A new review published in Environmental Monitoring and Assessment argues that the missing ingredient in managing these pollutants is integration. Conventional monitoring remains largely reactive: samples are collected periodically, shipped to centralized laboratories, and analyzed with chromatographic and spectrometric techniques that offer excellent sensitivity but are labor intensive, time consuming, expensive, and limited to discrete snapshots in space and time. Remediation technologies such as advanced oxidation, adsorption, electrochemical processes, and bioremediation have expanded considerably, but their effectiveness depends strongly on pollutant characteristics and operating conditions. The review’s central claim is that monitoring, prediction, and treatment have developed as separate technological components, and that connecting them through artificial intelligence, machine learning, Internet of Things sensor networks, cloud computing, and edge analytics could enable a closed-loop architecture in which environmental data continuously inform predictive models, remediation decisions, and system optimization.

The authors, who used a structured literature review informed by PRISMA principles covering publications from 2010 to 2026, first survey the analytical foundations of the field. Gas chromatography remains the benchmark for separating semi-volatile, nonpolar pollutants such as PCBs and polybrominated diphenyl ethers, while two-dimensional gas chromatography improves resolution for dioxins and PCBs in complex matrices. Liquid chromatography handles polar and thermally labile compounds including PFAS, and gas chromatography–mass spectrometry is described as the gold standard for sensitivity and specificity, with high-resolution and tandem mass spectrometry enabling precise quantification when combined with isotope dilution. These laboratory methods achieve detection limits at the femtogram-to-picogram level, but their cost is substantial: commercial laboratory pricing for gas chromatography–mass spectrometry ranges from 150 to 800 dollars per sample, with routine analysis typically between 150 and 300 dollars and turnaround times measured in days.

Biosensors offer a contrasting paradigm, converting specific binding events into optical, electrochemical, or thermal signals that can be read on site. The review catalogs striking examples: a quartz crystal microbalance immunosensor detecting PCB congeners at limits from 6 parts per billion to 1 part per million with responses under 20 minutes in aqueous solution; a surface plasmon resonance immunosensor detecting naphthalene in seawater at 0.76 parts per billion; a whole-cell bioluminescent biosensor using Sphingobium xenophagum for PBDEs at roughly 10 nanomolar; an enzyme-based electrochemical sensor for PCBs reaching 0.016 to 0.019 micrograms per liter; and a genetically engineered bacterial fluorescent biosensor for PFOA and PFOS with a detection limit of 10 nanograms per liter, validated in river water, wastewater, and drinking water. Aptamer-based and molecularly imprinted polymer sensors are highlighted as more robust alternatives to fragile antibodies and enzymes.

Yet the review is candid about biosensor limitations. Biological recognition elements are inherently sensitive to environmental stress: enzymes denature, antibodies degrade, and whole-cell systems lose metabolic activity, compromising reliability in the field. Matrix interference from organic matter, salts, and particulates causes non-specific binding, surface fouling, and signal suppression. Whole-cell biosensors, which uniquely report the bioavailable fraction of pollutants, require hours to a full day to respond because of metabolic activation and reporter gene expression, creating an inherent trade-off between sensitivity, speed, and specificity. Many devices are validated only with spiked or purified samples over short periods, while long-term monitoring under fluctuating pH, temperature, and biofouling remains rare.

On the computational side, the review documents machine learning’s expanding role. A hybrid ARIMA–LSTM model, validated with data from the Great Lakes Integrated Atmospheric Deposition Network, outperformed standalone models in forecasting pollutant concentrations and projected a general decline in POP levels over the next five years with seasonal fluctuations. A separate enhanced LSTM model achieved the lowest test error for Great Lakes concentrations and revealed that summer and fall levels ran roughly 40 percent higher than winter values. Quantitative structure–activity relationship modeling predicted air half-lives of 60 POPs with coefficients of determination around 0.94. Other studies traced groundwater contamination sources with a partial least squares discriminant model, predicted PCB concentrations in human breast milk with an explainable ExtraTrees model, and coupled fluorescence spectroscopy with a convolutional neural network to screen waste for POPs in about 10 minutes with 75 percent classification accuracy. In treatment, gradient boosting models predicted photocatalytic PFOA degradation and PFAS adsorption, while an optimized GBM model explained 97.5 percent of the variance in PFOS rejection by nanofiltration.

The remediation side shows equally notable laboratory results. A catalytic persulfate system using a manganese–cerium–alumina catalyst removed more than 85 percent of diverse POPs from landfill leachate, and an integrated system combining catalytic pre-oxidation with immobilized, genetically engineered microbes achieved over 96.8 percent removal while cutting operational costs and greenhouse gas emissions. Bio-electro-Fenton systems pairing microbial fuel cells with electrochemical Fenton reactions removed 98 percent of bisphenol-S at low concentrations while generating electrical power. A circular approach fermenting contaminated wetland plant biomass removed 87 to 95 percent of sequestered POPs and converted the waste into a pathogen-resistant liquid fertilizer. One fully integrated Internet of Things system in a subway station combined water quality sensing, threshold-triggered alerts, and an automated photocatalytic ozone reactor that degraded 93 percent of 2-chlorophenol, 86 percent of 2,4-dichlorophenol, and 83 percent of 2,4,6-trichlorophenol.

However, the review insists that these numbers should be read with caution. Laboratory-scale remediation rarely survives contact with field conditions, where landfill leachate chemical oxygen demand can range from 140 to 152,000 milligrams per liter and energy demands vary from about 0.37 to 39.2 kilowatt-hours per cubic meter depending on the process. Machine learning performance is strongly context dependent: random forests excelled for POP concentration prediction, gradient boosting for photocatalytic degradation, and LSTM networks for temporal forecasting, differences that reflect dataset structure rather than intrinsic algorithmic superiority. Data scarcity fundamentally limits generalizability, and biodegradation regression models achieved coefficients of determination of only 0.56 for sediments and 0.31 for sludge, with reasonable predictions for just 41 percent of more than 4,000 persistent chemicals. Only 22.5 percent of machine learning datasets assessed in Earth-observation applications reported any uncertainty information. Cybersecurity adds another layer of risk, with a reported 41 percent increase in weekly attacks on Internet of Things devices from 2022 to 2023, and biofouling can consume up to half of aquatic sensor operational budgets.

The review’s conclusion is that the future of persistent organic pollutant management lies in integration rather than technological substitution. Advanced analytical methods, continuous sensing, process-based knowledge, and remediation technologies should operate as complementary components of a validated framework, with artificial intelligence used first for screening, anomaly detection, and early warning while confirmatory regulatory determinations remain anchored to validated laboratory analysis. Emerging directions such as digital twins, federated learning, physics-informed machine learning, explainable artificial intelligence, and self-optimizing treatment systems could accelerate this shift, but only if paired with standardized, geographically diverse, FAIR-compliant datasets, harmonized calibration and benchmarking procedures, and independently validated pilot- and field-scale demonstrations. Until then, the gap between promising laboratory results and reliable real-world deployment remains the defining challenge for intelligent environmental protection.

Subject of Research: A review of artificial intelligence, machine learning, and Internet of Things technologies for monitoring, predicting, and remediating persistent organic pollutants

Article Title: From smart sensing to adaptive remediation: artificial intelligence and Internet of Things for persistent organic pollutants

Article References: Chawraba, K., Hussain, A., Alharbi, O. A., & Abdelhaleem, A. (2026). From smart sensing to adaptive remediation: artificial intelligence and Internet of Things for persistent organic pollutants. Environmental Monitoring and Assessment, 198(11), Article 1193. https://doi.org/10.1007/s10661-026-15975-x

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15975-x

Keywords: persistent organic pollutants, artificial intelligence, machine learning, Internet of Things, biosensors, environmental monitoring, remediation, advanced oxidation processes, Stockholm Convention, PFAS, PCBs, digital twins