聚变科学与工程资源库
检索物理代码、实验数据、代理模型、工程几何、技术文档与验证资料，并核对来源、版本和使用条件。
Dual-Stage Toxic Comment Detection System: Binary and Multi-Class Classification
ZenodoThis paper presents our latest evaluation on a unified toxic comment dataset (merged from the Jigsaw Toxic Comment Classification Chal- lenge and the Unintended Bias in Toxicity datasets), outlining both performance metrics and future enhancements. We detail the dataset merging process and quality checks, discuss evaluation results for our dual-stage model (bi- nary and multiclass classification), incorporate an advanced deep neural network (DNN) ap- proach with semantic embeddings and engi- neered features, and propose steps to expand our system to address dialect and demographic fairness.
Eye-tracking dataset: Seeing without reporting (visual attention to vulnerable individuals)
Zenodo已收录公开题录与来源记录，正文内容及使用条件请通过原始发布页面核对。
Simulation Results of a Nitrogen Rejection Unit
Tecnológico de MonterreyThe details regarding the data description are provided in the README file. The data corresponding to Figure 2 were obtained from a gas chromatography analysis of the gas plant. The data corresponding to Figures 3, 5, 9, and 10 were recorded in real time using the Aspen HYSYS® simulation environment. The data are provided in raw form to facilitate reproducibility of the results presented in the submitted manuscript.
Bone strength analyzer and monitoring device for lower limb external fixation
Universiti Malaysia Pahang Al-Sultan AbdullahThe procedure for external fixator removal in lower limb fractures is typically based on radiographic data, subjects’ patients to ionizing radiation, and provides minimal real-time information about the healing process. This project suggests a sensor-based Internet-of-Things apparatus, which will measure bone strength during the recuperating process based on load cells, HX711 amplifiers, and a Wemos ESP8266 microcontroller. The system provides real-time feedback in the form of LED indicators and a buzzer, whereas remote monitoring is supported by the Blynk dashboard. Measurement accuracy of over 90% was carried out as per the experimental validation conducted under the simulation of various loads (3-9 kg) and with clear stage indicators of critical, partial, and full recovery. The device offers continuous monitoring and objective bone healing evaluation with no radiation in comparison to conventional imaging. The limitation of the study is the limited range of loads in prototype testing, which could affect accuracy in particular situations. However, the results represent the possible clinical relevance of introducing real-time biomechanical surveillance into the process of fracture treatment, hence contributing to safer rehabilitation and more reasonable decisions related to the fixator removal.
Sampling Effort Data
ZenodoOpen access data on Mosquito Alert participation and sampling effort. This dataset contains data on participation and sampling effort in the Mosquito Alert citizen science system. It can be used for a variety of purposes, including (a) to adjust estimates of mosquito population densities and human-mosquito encounters based on sampling effort, and (b) to better understand the dynamics of citizen scientists' participation. The data is organized spatially by grids of "sampling cells," drawn at intervals of 0.05 degree and 0.025 degree latitude and longitude, and it is based on optional anonymous background tracks from the Mosquito Alert app. The dataset includes raw track counts aggregated in sampling cells, along with estimates of sampling effort based on a model of participants' propensity to send any report as a function of the time elapsed since they first began participating. The repository is hosted on both Zenodo and GitHub and contains the following files: sampling_effort_daily_cellres_025.csv.gz - Daily participation and sampling effort in 0.025 degree sampling cells. sampling_effort_daily_cellres_025_metadata.json - Metadata for the 0.025 degree sampling cell data. sampling_effort_daily_cellres_05.csv.gz - daily participation and sampling effort in 0.05 degree sampling cells. sampling_effort_daily_cellres_05_metadata.json - Metadata for the 0.05 degree sampling cell data. CITATION.cff - Shows how to cite this dataset. LICENSE - License for this dataset. .gitignore - Specifies which files are excluded from the git repository. README.md - This file. .github/workflows/auto-release.yml - Script used for automating releases from GitHub.
Graph neural network-based biomedical misinformation detection with semantic consistency analysis
SRM University - Ramapuram CampusConventional misinformation detection approaches primarily rely on textual features and deep learning (DL) classifiers, which often fail to capture complex relationships among biomedical entities and the underlying scientific context of health claims. To address this limitation, this study proposes a graph neural network (GNN)-based biomedical misinformation detection framework that integrates knowledge graph propagation with semantic consistency verification. Initially, key biomedical entities such as diseases, treatments, and biological processes are extracted and mapped into a structured biomedical knowledge graph (BKG) to represent semantic relationships. A graph attention network (GAT) is then employed to model relational dependencies and propagate contextual information across connected entities, enabling the detection of hidden inconsistencies in biomedical claims. The proposed model is evaluated using benchmark biomedical misinformation datasets, including Reliable COVID-19 News Dataset, 2021 (ReCOVery), COVID-19 Healthcare Misinformation Dataset, 2020 (CoAID), and 2018–2020 biomedical health news corpus (HealthStory). Experimental results demonstrate that the proposed framework achieves an average detection accuracy of 96.3%, outperforming conventional long shortterm memory (LSTM), convolutional neural networks (CNN), and transformer-based models in terms of precision, recall, and F1-score. The findings highlight that integrating structured biomedical knowledge with graph-based reasoning significantly enhances the reliability and interpretability of misinformation detection systems.
Code for: Phenological Fluctuations Induce Seasonal Asymmetric Warming in Boreal Forests
ZenodoCode accompanying the manuscript "Phenological Fluctuations Induce Seasonal Asymmetric Warming in Boreal Forests" (Nature Communications, 2026). The repository contains the Google Earth Engine JavaScript scripts for seasonal compositing of the MODIS products, quality masking, the persistent-forest mask and computation of the leaf area change and change-rate layers, and the Python and R scripts for the paired-grid matching, the inverse distance weighting climate-background removal, the moving-window Huber regression, the asymmetry index, the structural equation modelling and the robustness analyses (block bootstrap, baseline stratification and truncation sensitivity). All input data are public MODIS products and the FLUXNET2015 dataset, as listed in the Data Availability statement of the paper.
Sparse Quotient-Threshold Transform (SQTT): benchmark data, reference code, and reproducibility package
University of PhayaoThis version 1.0.0 research compendium supports the computational evaluation of the Sparse Quotient-Threshold Transform (SQTT), an exact method for computing all outputs of a knapsack-threshold transform and its additive generalized-assignment extension. The compendium contains the MATLAB reference implementation, declared exact comparators, instance generators, correctness checks, the frozen primary experimental protocol, analysis scripts, preservation-friendly data tables, figure-source data, rendered figures, and SHA-256 artifact manifests. The empirical data comprise 1,788 completed full-benchmark case-method trials over 610 generated cases and a balanced 30-run scalar scalability panel for k = 18 and k = 20. The balanced panel is a post-inspection exploratory supplement and is not part of the prespecified primary benchmark. Eight accepted k = 22 observations are supplied in a separate exploratory resource-boundary dataset: three completed outcomes and five verified memory-limit outcomes. These k = 22 observations are not included in the primary runtime comparisons, and no missing runtime is imputed. The statistical unit is the generated case rather than an individual timing repeat. Timing repeats are summarized within each case. Paired log-speed ratios are defined as ln(baseline time / SQTT time), and uncertainty is reported using stratified case-level percentile bootstrap 95% confidence intervals. The archive includes the derived case-level dataset, summary tables, performance-profile curves, scaling summaries, resource-boundary summaries, and an independent correctness gate covering 2,000 scalar and 500 generalized-assignment cases with zero failures. Code is released under the MIT License. Data, tables, figures, protocols, and documentation are released under Creative Commons Attribution 4.0 International. File-level provenance and SHA-256 checksums are provided in the package manifest.
Vision-Language Models for Low-Vision Navigation Assistance in Cluttered Environments
Birla Institute of Technology and Science, PilaniNavigating busy, unstructured roads is a daily challenge for people with visual impairments, while many assistive systems do not explicitly account for the clutter and heterogeneous traffic found in Indian street environments. We present an end-to-end vision-to-audio navigation pipeline that transforms monocular video frames into priority-ordered spoken directional commands. The reported system combines a YOLOv8n detector fine-tuned on the Indian Driving Dataset (IDD), MiDaS DPT-LeViT monocular depth estimation, DeepSORT tracking, field-of-view-aware spatial reasoning, temporal smoothing, and an eight-layer rule-based navigation planner. The planner maps perceived hazards to concise commands such as STOP, AVOID, MOVE LEFT/RIGHT, and CONTINUE FORWARD, while a smart text-to-speech gate suppresses passive descriptions when an urgent command is pending. An optional BLIP-based visual validation stage is used in image mode. On the reported 60-frame manually annotated Indian-street evaluation set, the complete pipeline achieves 97.9% hazard recall, 92.8% distance accuracy, 0.407 walkable-corridor IoU, and 84.6 ms average latency per frame. On the IDD detection test set, fine-tuning improves mAP@50 from 0.389 for the COCO- pretrained baseline to 0.542, with mAP@50–95 increasing from 0.224 to 0.351. Ablations show that field-of-view-aware depth calibration is critical for proximity estimation, while temporal tracking is important for instruction stability and motion-aware warnings. The system is deliberately modular: the perception and spatial reasoning stack supplies grounded state information, while deterministic planning converts that state into actionable audio rather than relying on unconstrained generative narration.