Fant 9999 publikasjoner. Viser side 4 av 400:
2025
This study builds upon the findings of a FAIRMODE intercomparison exercise conducted in a district of Antwerp, Belgium, where a comprehensive dataset of air pollutant measurements (air quality stations and passive samplers) was available. Long-term average NO2 concentrations at very high spatial resolution were estimated by several dispersion modelling systems (Martín et al., 2024) to investigate the ability of these to capture the detailed spatial distribution of NO2 concentrations at the microscale in urban environments. In this follow-up research, we extend the analysis by evaluating the capability of these modelling systems to predict the NO2 annual limit value exceedance areas (LVEAs) and spatial representativeness areas (SRAs) for NO₂ at two reference air quality stations. The different modelling approaches used are based on CFD, Lagrangian, Gaussian, and AI-driven models.
The different modelling approaches are generally good at predicting the LVEA and SRAs of urban air quality stations, although a small SRA (corresponding to low concentration tolerances or the traffic station) is more difficult to predict correctly. However, there are notable differences in performance among the modelling systems. Those based on CFD models seem to provide more consistent results predicting LVEAs and SRAs. Then, lower accuracy is obtained with AI-based systems, Lagrangian models, and Gaussian models with street canyon parameterizations. The Gaussian models with street-canyon parametrizations show significantly better results than models using simply a Gaussian dispersion parametrization.
Furthermore, little differences are observed in most of the statistical indicators corresponding to the LVEA and SRA estimates obtained from the unsteady full month CFD simulations compared to those from the scenario-based CFD simulation methodologies, but there are some noticeable differences in the LVEA or SRA (traffic station, 10 % tolerance) sizes. The number of scenarios does not seem to be relevant to the results. Different bias correction methodologies are explored.
2025
2025
Sb-PiPLU: A Novel Parametric Activation Function for Deep Learning
The choice of activation function—particularly non-linear ones—plays a vital role in enhancing the classification performance of deep neural networks. In recent years, a variety of non-linear activation functions have been proposed. However, many of these suffer from drawbacks that limit the effectiveness of deep learning models. Common issues include the dying neuron problem, bias shift, gradient explosion, and vanishing gradients. To address these challenges, we introduce a new activation function: Softsign-based Piecewise Parametric Linear Unit (Sb-PiPLU). This function offers improved non-linear approximation capabilities for neural networks. Its piecewise, parametric design allows for greater adaptability and flexibility, which in turn enhances overall model performance. We evaluated Sb-PiPLU through a series of image classification experiments across various Convolutional Neural Network (CNN) architectures. Additionally, we assessed its memory usage and computational cost, demonstrating that Sb-PiPLU is both stable and efficient in practical applications. Our experimental results show that Sb-PiPLU consistently outperforms conventional activation functions in both classification accuracy and computational efficiency. It achieved higher accuracy on multiple benchmark datasets, including CIFAR-10, CINIC-10, MWD, Brain Tumor, and SVHN, surpassing widely-used functions such as ReLU and Tanh. Due to its flexibility and robustness, Sb-PiPLU is particularly well-suited for complex image classification tasks.
2025
2025
CompSafeNano project: NanoInformatics approaches for safe-by-design nanomaterials
The CompSafeNano project, a Research and Innovation Staff Exchange (RISE) project funded under the European Union's Horizon 2020 program, aims to advance the safety and innovation potential of nanomaterials (NMs) by integrating cutting-edge nanoinformatics, computational modelling, and predictive toxicology to enable design of safer NMs at the earliest stage of materials development. The project leverages Safe-by-Design (SbD) principles to ensure the development of inherently safer NMs, enhancing both regulatory compliance and international collaboration. By building on established nanoinformatics frameworks, such as those developed in the H2020-funded projects NanoSolveIT and NanoCommons, CompSafeNano addresses critical challenges in nanosafety through development and integration of innovative methodologies, including advanced in vitro models, in silico approaches including machine learning (ML) and artificial intelligence (AI)-driven predictive models and 1st-principles computational modelling of NMs properties, interactions and effects on living systems. Significant progress has been made in generating atomistic and quantum-mechanical descriptors for various NMs, evaluating their interactions with biological systems (from small molecules or metabolites, to proteins, cells, organisms, animals, humans and ecosystems), and in developing predictive models for NMs risk assessment. The CompSafeNano project has also focused on implementing and further standardising data reporting templates and enhancing data management practices, ensuring adherence to the FAIR (Findable, Accessible, Interoperable, Reusable) data principles. Despite challenges, such as limited regulatory acceptance of New Approach Methodologies (NAMs) currently, which has implications for predictive nanosafety assessment, CompSafeNano has successfully developed tools and models that are integral to the safety evaluation of NMs, and that enable the extensive datasets on NMs safety to be utilised for the re-design of NMs that are inherently safer, including through prediction of the acquired biomolecule coronas which provide the biological or environmental identities to NMs, promoting their sustainable use in diverse applications. Future efforts will concentrate on further refining these models, expanding the NanoPharos Database, and working with regulatory stakeholders thereby fostering the widespread adoption of SbD practices across the nanotechnology sector. CompSafeNano's integrative approach, multidisciplinary collaboration and extensive stakeholder engagement, position the project as a critical driver of innovation in NMs SbD methodologies and in the development and implementation of computational nanosafety.
2025
2025
2025
Citizen-operated low-cost air quality sensors (LCSs) have expanded air quality monitoring through community engagement. However, still challenges related to lack of semantic standards, data quality, and interoperability hinder their integration into official air quality assessments, management, and research. Here, we introduce FILTER, a geospatially scalable framework designed to unify, correct, and enhance the reliability of crowd-sourced PM2.5 data across various LCS networks. FILTER assesses data quality through five steps: range check, constant value detection, outlier detection, spatial correlation, and spatial similarity. Using official data, we modeled PM2.5 spatial correlation and similarity (Euclidean distance) as functions of geographic distance as benchmarks for evaluating whether LCS measurements are sufficiently correlated/consistent with neighbors. Our study suggests a −10 to 10 Median Absolute Deviation threshold for outlier flagging (360 h). We find higher PM2.5 spatial correlation in DJF compared to JJA across Europe while lower PM2.5 similarity in DJF compared to JJA. We observe seasonal variability in the maximum possible distance between sensors and reference stations for in-situ (remote) PM2.5 data correction, with optimal thresholds of ∼11.5 km (DJF), ∼12.7 km (MAM), ∼20 km (JJA), and ∼17 km (SON). The values implicitly reflect the spatial representativeness of stations. ±15 km relaxation for each season remains feasible when data loss is a concern. We demonstrate and validate FILTER's effectiveness using European-scale data originating from the two community-based monitoring networks, sensor.community and PurpleAir with QC-ed/corrected output including 37,085 locations and 521,115,762 hourly timestamps. Results facilitate uptake and adoption of crowd-sourced LCS data in regulatory applications.
2025
2025