This study examined the influence of biomass burning on the variability of fine particulate matter (PM2.5, aerodynamic diameter ≤ 2.5 µm) at a suburban receptor site in Nakhon Pathom, Thailand, during the dry seasons of 2023 and 2024. Regional transport was characterized using backward trajectory clustering and concentration-weighted trajectory analysis, fire characteristics using satellite-derived fire radiative power, and local source–receptor relationships using time-lagged Spearman correlation stratified by wind direction. Northeasterly transport dominated in both seasons, declining from 57.43
Artificial intelligence (AI) has become increasingly important in atmospheric aerosol and particulate-matter (PM) research because it can model nonlinear relationships, integrate heterogeneous observations, and generate high-resolution spatiotemporal information. This study maps the evolution and knowledge structure of AI-based atmospheric aerosol and PM research using 4,872 Web of Science (WoS) publications published through December 31, 2025. Bibliometric and text-mining analyses were used to characterize publication trends, leading countries and institutions, major journals, citation patterns, keyword co-occurrence networks, and abstract-based thematic clusters. In addition, 99 publications indexed in the Korea Citation Index (KCI) were examined separately to provide supplementary context on temporal publication trends in the Korean domestic literature. WoS publication activity increased sharply after 2019, with China contributing the largest publication output and total citation count, while several lower-output countries showed higher mean citations per publication. Keyword-network analysis identified machine learning as a central theme linking PM prediction, deep learning, aerosol optical depth, exposure assessment, and satellite remote sensing. Abstract-based clustering revealed five major thematic domains spanning prediction modeling, PM concentration and exposure, satellite-based aerosol retrieval, emission-related analysis, and aerosol monitoring and environmental applications. The results indicate a shift from model-specific prediction toward multisource integration, spatiotemporal analysis, remote sensing, and exposure-oriented applications. These findings provide a structured overview of the field and highlight the need for more interpretable, uncertainty-aware, and transferable AI approaches.
Sick Building Syndrome (SBS) is a widespread occupational health issue traditionally diagnosed through reactive, subjective symptom questionnaires. This study addresses the critical need for real-time, objective mitigation strategies by developing an automated early warning system to predict multi-categorical SBS risk using machine learning. We conducted a cross-sectional epidemiological study in a Malaysian educational institution, securing a unified dataset from 513 respondents across ten classrooms. Real-time physical environmental parameters and aerosol contaminants, specifically Particulate Matter (PM2.5) and Total Volatile Organic Compounds (TVOC), were captured at a seated breathing zone of 110 cm using high-precision occupational hygiene instruments. This environmental data was synchronized with self-reported health outcomes from a revised MM040NA questionnaire. To address the naturally low prevalence of severe symptoms and mitigate overfitting, our predictive modelling pipeline incorporated a 10-fold stratified cross-validation framework alongside the Synthetic Minority Over-sampling Technique (SMOTE) applied exclusively to training folds. We evaluated five classifiers: Artificial Neural Networks (ANN), Random Forests, Support Vector Classifiers (SVC), K-Nearest Neighbours (KNN), and Logistic Regression (LR), optimizing probability thresholds via the F2-score to prioritize recall. The results demonstrate that General SBS symptoms can be predicted with exceptional sensitivity, with ANN, SVC, and LR models achieving recall rates exceeding 0.90. For predicting the aggregate risk of Total SBS, the distance-based KNN algorithm produced the most balanced F1-score (0.419), serving as a theoretical proof-of-concept for risk clustering. Finally, this research confirms the viability of deploying resource-efficient linear or neural network models onto resource-constrained microcontrollers, providing concrete memory and latency benchmarks for a scalable Internet of Things (IoT) architecture designed for proactive indoor environmental health monitoring.
This study designed an improved parallel-plate particle separator, the Keio Measurement System of Aerosol Charging State (K-MACS), using numerical simulations to achieve a stable laminar flow field in practical devices. Combined with optical particle counters (OPCs), the system enables high-temporal-resolution (20 s) measurements of number concentrations of positively charged, uncharged, and negatively charged particles, allowing estimation of charge distributions for submicron particles (0.3–0.5 µm). The average particle charge number (pave) and standard deviation (1σ) were obtained by fitting the distributions to a Gaussian function. Experimental evaluation demonstrated that multiply charged particles moved toward oppositely charged electrodes and identified an optimum voltage of approximately 5.4 kV, where experimental and theoretical charge distributions agreed well. Measurement uncertainties for 1σ and distribution amplitude were reduced to 5
This study characterizes the chemical and optical properties of carbonaceous aerosols in urban Chiang Mai, Thailand, during an intensive monitoring period in March 2024. The campaign coincided with a strong El Niño event, resulting in extreme high temperatures and severe drought conditions. The mean PM2.5 concentration was 64.3 µg/m³, exceeding the WHO 24-hour guideline (15 µg/m³) by approximately 4.3 times. High OC/EC ratios (mean 6.0) and a strong correlation with CO (R = 0.71, 924.5 ppb) indicated a strong influence of biomass burning on particulate air quality, while urban baseline emissions also contributed. Furthermore, severe nighttime atmospheric stagnation trapped pollutants, exacerbating high-concentration episodes. Source apportionment via an Aethalometer model revealed that biomass burning-derived black carbon (BCbb) drove the overall variability during these episodes. Notably, brown carbon (BrC) significantly enhanced shortwave light absorption at 370 nm, accounting for a campaign mean contribution of 35.1
