Timea Bezdan

Benzene is a carcinogenic urban volatile organic compound whose variability reflects interactions among emissions, atmospheric chemistry, and meteorology. We quantified relationships between hourly benzene concentrations and O₃, CO, SO₂, NO₂, and more than twenty meteorological predictors using regulatory air-quality observations from the Ksaverska cesta urban background station in Zagreb, Croatia, complemented by NOAA GDAS meteorological data (2017–2023; n = 39,124). Metaheuristically optimized tree-ensemble models were interpreted using iterative Shapley Additive Global Importance (iSAGE), SHapley Additive exPlanations (SHAP), and an environmental settings framework. The novelty of the study lies in integrating global predictor importance, signed local SHAP attributions, and environmental-setting classification to interpret benzene variability within recurring emission–chemistry–meteorology regimes. The best-performing Extra Trees model achieved R² = 0.87, RMSE = 0.46 µg m⁻³, and MAE = 0.24 µg m⁻³. iSAGE identified CO (1.47), O₃ (0.55), and near-surface air temperature (0.35) as the leading predictors of modeled benzene concentrations, whereas SO₂ (0.18) and NO₂ (0.07) had substantially lower global importance. SHAP showed a positive CO-benzene association, with relative contributions shifting from negative to positive at approximately 0.3–0.6 mg m⁻³; O₃ contributions became predominantly negative above 30–40 µg m⁻³; and temperature contributions shifted from positive below 7 °C to negative across 7–12 °C. Environmental setting analysis associated the strongest benzene accumulation with stable winter conditions and the weakest accumulation with well-ventilated, photochemically active regimes. These site- and model-specific results reveal nonlinear co-pollutant and meteorological influences on urban benzene variability. Theoretically, the study advances the interpretation of urban VOC variability by linking model-based predictor attributions with physically interpretable atmospheric regimes; practically, it supports the identification of conditions favoring benzene accumulation and can inform regime-sensitive air-quality assessment and management.

Benzene is a major carcinogenic urban pollutant whose variability reflects interactions between emission sources, human activity, and atmospheric conditions. Although COVID-19 restrictions generally reduced traffic-related emissions, the combined effects of mobility changes, residential activity, and policy interventions on urban benzene dynamics remain insufficiently understood. This study investigated benzene variability in Belgrade, Serbia, during the COVID-19 state of emergency, from 15 March to 6 May 2020. A multi-source dataset was used, integrating high-resolution VOC measurements by PTR-quad-MS, meteorological variables, regulatory air-quality indicators, epidemiological data, mobility proxies, and quantified government-response measures. Tree-based ensemble machine-learning models, metaheuristic hyperparameter optimization, and explainable artificial intelligence methods, including SAGE and SHAP, were applied to examine non-linear and time-lagged relationships within the urban atmospheric system. The results showed that benzene variability was primarily associated with co-measured non-target VOCs, reflecting shared urban emission-source structures. Mobility and policy-related predictors contributed through short delayed responses, with an estimated response window of approximately 48–72 h. Sustained mobility reductions were associated with lower benzene concentrations, whereas increased residential activity partially offset traffic-related reductions. Within the Belgrade case study, these findings demonstrate the potential of interpretable machine learning to extract robust patterns from heterogeneous urban environmental datasets, while emphasizing the need for validation across additional cities and non-pandemic conditions before broader generalization.

Polychlorinated biphenyls (PCBs) remain a major concern in marine ecosystems, where bioaccumulation in fish occurs as complex congener mixtures whose dynamics challenge conventional indicator approaches. This study develops and evaluates a data-driven framework for refining mixture-based indicators of PCB contamination by integrating ensemble machine learning with explainable artificial intelligence. Focusing on PCB-138 as a target indicator of cumulative PCB burden, we analyse concentrations of 24 organochlorines together with biological covariates in four Mediterranean edible pelagic fish species (sardine, anchovy, horse mackerel, and chub mackerel). Comparative evaluation of indicator performance shows that alternative congener combinations, including i4 PCBs (-138, -153, -170, -180), i6 PCBs (-138, -153, -170, -180, -118, -123), and mixtures incorporating DDD and DDE, more effectively represent total PCB burden than traditional indicator groups. Clustering identifies two distinct bioaccumulation settings, characterized by high-concentration coherent congener effects and low-concentration heterogeneous responses, demonstrating that indicator performance depends on concentration range and mixture context. The study illustrates how interpretable machine learning approaches can serve as formal tools for indicator evaluation and optimisation, strengthening long-term monitoring and management of legacy contaminants in marine ecosystems, particularly under conditions of persistent exposure and renewed inputs from sediment remobilization and riverine transport.

We investigated benzene variability in an urban environment using an interpretable, setting-based artificial intelligence framework. A seven-year dataset (2017–2023) of hourly pollutant concentrations (benzene, NO₂, SO₂, CO, O₃) measured in Zagreb (Croatia) was analyzed, as were meteorological variables. Multiple-ensemble decision tree models were developed, with hyperparameters optimized using metaheuristic algorithms. The best-performing model, Extra Trees optimized by the Sine Cosine Algorithm, achieved an R² of 0.87. Model interpretation employed Shapley additive explanations (SHAP), followed by PaCMAP embedding and HDBSCAN clustering to identify coherent environmental settings. Seven settings (C0–C6) and one residual group were identified, representing pollution-enhancing, suppressing, and transitional regimes. Two settings dominated benzene extremes. C6 reflected winter stagnation, characterized by strong combustion influence (CO contribution of 11.9%), shallow boundary layers (~290 m), weak winds, and high humidity. C4 represented a synoptic stability regime with enhanced heat fluxes and diminished after the COVID-19 period, consistent with altered anthropogenic activity. Low-benzene settings (C0, C1, C3) were associated with stronger mixing and higher oxidizing capacity, while transitional settings (C2, C5) reflected moderate conditions. Overall, the results show that a small number of environmental settings governed the benzene extremes, providing a transferable and interpretable framework for air quality assessment and policy support.

Human milk has been used for over 70 years to monitor pollutants such as polychlorinated biphenyls (PCBs) and organochlorine pesticides (OCPs). Despite the growing body of data, our understanding of the pollutant exposome, particularly co-exposure patterns and their interactions, remains limited. Artificial intelligence (AI) offers considerable potential to enhance biomonitoring efforts through advanced data modelling, yet its application to pollutant dynamics in complex biological matrices such as human milk remains underutilized. This study applied an AI-based framework, integrating machine learning, metaheuristic hyperparameter optimization, explainable AI, and postprocessing, to analyze PCB-170 levels in breast milk samples from 186 mothers in Zadar, Croatia. Among 24 analyzed POPs, the most influential predictors of PCB-170 concentrations were hexa- and hepta-chlorinated PCBs (PCB-180, -153, and -138), alongside p,p’-DDE. Maternal age and other POPs exhibited negligible global influence. SHAP-based interaction analysis revealed pronounced co-behavior among highly chlorinated congeners, especially PCB-138–PCB-153, PCB-138–PCB-180, and PCB-180–PCB-153. These findings highlight the importance of examining pollutant interactions rather than individual contributions alone. They also advocate for the revision of current monitoring strategies to prioritize multi-pollutant assessment and focus on toxicologically relevant PCB groups, improving risk evaluation in real-world exposure scenarios.

The world's ever-increasing need for energy demands innovation and resourcefulness from engineers and power companies. Reducing transportation losses and providing reliable power are just some of the challenges facing power production from renewable resources. Nevertheless, addressing these challenges is essential for improving sustainability and slowing down climate change. The sporadic nature of power production from solar panels makes it difficult to integrate into existing power grid infrastructure. A robust forecasting system could help mitigate these issues. This study investigates the application of long short term memory (LSTM)-based neural networks to tackle production forecasting. Since forecasting models rely on proper hyperparameter selection, this study presents a modified optimizer for hyperparameter tuning to improve performance. Simulations on real world data have shown promising results, with LSTM based networks optimized by a modified metaheuristic based on the PSO showcasing favorable outcomes with a mean absolute percentage error (MAPE) as low as 4.112%.

This study introduces a novel artificial intelligence (AI) modeling framework that combines machine learning algorithms optimized through metaheuristics with explainable AI to capture complex interactions among pollutant concentrations, meteorological data, and socio-economic indicators. Applied to a COVID-19-related dataset comprising 404 variables, with benzene concentrations as the target—measured using proton transfer reaction–mass spectrometry in Belgrade, Serbia—the framework demonstrated exceptional sensitivity in assessing the impact of complex environmental and societal changes during the pandemic. Explainable AI techniques, such as SHAP and SAGE, were employed to reveal the influence of each predictor, while the clustering of SHAP values identified distinct environmental settings that influenced benzene behavior. Three distinct settings were identified regarding benzene levels during the onset of the state of emergency. The first, involving local petroleum-related activities, biomass burning, chemical manufacturing, and traffic, led to a 15.7% reduction in benzene levels. The second, characterized by non-combustion processes, nocturnal chemistry, and the specific meteorological context, resulted in a 51.9% increase. The third, driven by local industrial processes, contributed to a modest 2.33% reduction. The study underscored the critical role of environmental settings in shaping air pollutant behavior, emphasizing the importance of integrating broader environmental contexts into models to gain a more comprehensive understanding of air pollutants and their dynamics.

Currently, there are many different computer vision models available, and each has its unique characteristics

This study aims to investigate the impact of indoor sources and outdoor air on indoor PM_2.5-bound benzo(a)pyrene, with a focus on identifying emission sources and understanding the influence of environmental variables. For this purpose, we collected indoor and outdoor data on PM_2.5-bound PAHs, inorganic gaseous pollutants, trace metals, ions, radon, and meteorological parameters, resulting in a comprehensive dataset of 100 variables from an urban site in Belgrade, Serbia. We applied seven regression tree ensemble algorithms to interrelate the variables alongside six metaheuristic optimization algorithms to refine model accuracy and robustness. Subsequently, we explained the best-performing model locally using Shapley additive explanations and clustered variables with similar impacts into distinct groups. These groups were systematically characterized, defining them as environmental settings that shape benzo(a)pyrene dynamics. The setting resulting in the highest indoor benzo(a)pyrene concentrations (197% to 297% relative to the expected value) was dominated by outdoor emissions associated with residential heating and traffic (up to 140%) and indoor source identified as cooking. This integrated approach uniquely enables a quantitative assessment of the contributions from both indoor and outdoor emission sources to pollutant concentrations in indoor spaces, underscoring the importance of both in shaping indoor air quality. Unlike traditional source apportionment methods that assume linear source mixing, our approach integrates nonlinear interactions and contextual variables, such as meteorological conditions and outdoor pollutants, to better capture indoor air quality dynamics. The results also highlight the need for further studies to explore broader contextual factors and refine source attribution in complex urban settings.