Fant 2690 publikasjoner. Viser side 5 av 269:
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
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
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
Methane emissions from the Nord Stream subsea pipeline leaks
The amount of methane released to the atmosphere from the Nord Stream subsea pipeline leaks remains uncertain, as reflected in a wide range of estimates1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18. A lack of information regarding the temporal variation in atmospheric emissions has made it challenging to reconcile pipeline volumetric (bottom-up) estimates1,2,3,4,5,6,7,8 with measurement-based (top-down) estimates8,9,10,11,12,13,14,15,16,17,18. Here we simulate pipeline rupture emission rates and integrate these with methane dissolution and sea-surface outgassing estimates9,10 to model the evolution of atmospheric emissions from the leaks. We verify our modelled atmospheric emissions by comparing them with top-down point-in-time emission-rate estimates and cumulative emission estimates derived from airborne11, satellite8,12,13,14 and tall tower data. We obtain consistency between our modelled atmospheric emissions and top-down estimates and find that 465 ± 20 thousand metric tons of methane were emitted to the atmosphere. Although, to our knowledge, this represents the largest recorded amount of methane released from a single transient event, it is equivalent to 0.1% of anthropogenic methane emissions for 2022. The impact of the leaks on the global atmospheric methane budget brings into focus the numerous other anthropogenic methane sources that require mitigation globally. Our analysis demonstrates that diverse, complementary measurement approaches are needed to quantify methane emissions in support of the Global Methane Pledge19.
2025
2025
Aerosol hygroscopicity influenced by seasonal chemical composition variations in the Arctic region
In this study, we quantified aerosol hygroscopicity parameter using aerosol microphysical observation data (κphy), analyzing monthly and seasonal trends in κphy by correlating it with aerosol chemical composition over 6 years from April 2007 to March 2013 at the Zeppelin Observatory in Svalbard, Arctic region. The monthly mean κphy value exhibited distinct seasonal variations, remaining high from winter to spring, reaching its minimum in summer, followed by an increase in fall, and maintaining elevated levels in winter. To verify the reliability of κphy, we employed the hygroscopicity parameter calculated from chemical composition data (κchem). The chemical composition and PM2.5 mass concentration required to calculate κchem was obtained through Modern-Era Retrospective Analysis for Research and Applications, version 2 (MERRA-2) reanalysis data and the calculation of κchem assumed that Arctic aerosols comprise only five species: black carbon (BC), organic matter (OM), ammonium sulfate (AS), sea salt aerosol less than a diameter of 2.5 μm (SSA2.5), and dust aerosol less than a diameter of 2.5 μm (Dust2.5). The κchem had no distinct correlation but had a similar seasonal trend compared to κphy. The κchem value followed a trend of SSA2.5 and was much higher by a factor of 1.6 ± 0.3 than κphy on average, due to a large proportion of SSA2.5 mass concentration in MERRA-2 reanalysis data. This may be due to the overestimation of sea salt aerosols in MERRA-2 reanalysis. The relationship between monthly mean κphy and the chemical composition used to calculate κchem was also analyzed. The elevated κphy from October to February resulted from the dominant influence of SSA2.5, while the maximum κphy in March was concurrently influenced by increasing AS and Dust2.5 associated with long-range transport from mid-latitude regions during Arctic haze periods and by SSA mass concentration obtained from in-situ sampling, which remained high from the preceding winter. The relatively low κphy from April to September can be attributed to low SSA2.5 and the dominance of organic compounds in the Arctic summer. Either natural sources such as those of marine and terrestrial biogenic origin or long-range-transported aerosols may contribute to the increase in organic aerosols in summer, potentially influencing the reduction in κphy of atmospheric aerosols. To our knowledge, this is the first study to analyze the monthly and seasonal variation of aerosol hygroscopicity calculated using long-term microphysical data, and this result provides evidence that changes in monthly and seasonal hygroscopicity variation occur depending on chemical composition.
2025
Australia has significant sources of atmospheric methane (CH₄), driven by extensive coal and natural gas production, livestock, and large-scale fires. Accurate quantification and characterization of CH₄ emissions are critical for effective climate mitigation strategies in Australia. In this study, we employed an inverse analysis of atmospheric CH₄ observations from the GOSAT satellite and surface measurements from 2016 to 2021 to assess CH₄ emissions in Australia. The inversion process integrates anthropogenic and natural emissions as prior estimates, optimizing them with the NIES-TM-FLEXPART-variational model (NTFVAR) at a resolution of up to 0.1° × 0.1°. We validated the performance of our inverse model using data obtained from the United Nations Environment Program Methane Science (UNEP), Airborne Research Australia 2018 aircraft-based atmospheric CH₄ measurement campaigns. Compared to prior emission estimates, optimized emissions dramatically enhanced the accuracy of modeled concentrations, aligning them much better with observations. Our results indicate that the estimated inland CH4 emissions in Australia amount to 6.84 ± 0.51 Tg CH4 yr−1 and anthropogenic emissions amount to 4.20 ± 0.08 Tg CH4 yr−1, both slightly lower than the values reported in existing inventories. Moreover, our results unveil noteworthy spatiotemporal characteristics, such as upward corrections during the warm season, particularly in Southeastern Australia. During the three most severe months of the 2019–2020 bushfire season, emissions from biomass burning surged by 0.68 Tg, constituting over 71% of the total emission increase. These results highlight the importance of continuous observation and analysis of sectoral emissions, particularly near major sources, to guide targeted emission reduction strategies. The spatiotemporal characteristics identified in this study underscore the need for adaptive and region-specific approaches to CH₄ emission management in Australia.
2025
Potato plant disease detection: leveraging hybrid deep learning models
Agriculture, a crucial sector for global economic development and sustainable food production, faces significant challenges in detecting and managing crop diseases. These diseases can greatly impact yield and productivity, making early and accurate detection vital, especially in staple crops like potatoes. Traditional manual methods, as well as some existing machine learning and deep learning techniques, often lack accuracy and generalizability due to factors such as variability in real-world conditions. This study proposes a novel approach to improve potato plant disease detection and identification using a hybrid deep-learning model, EfficientNetV2B3+ViT. This model combines the strengths of a Convolutional Neural Network - EfficientNetV2B3 and a Vision Transformer (ViT). It has been trained on a diverse potato leaf image dataset, the “Potato Leaf Disease Dataset”, which reflects real-world agricultural conditions. The proposed model achieved an accuracy of 85.06, representing an 11.43 improvement over the results of the previous study. These results highlight the effectiveness of the hybrid model in complex agricultural settings and its potential to improve potato plant disease detection and identification.
2025