让变化经过筛选，
再抵达你。

研究者提出 SW-CraftNet，并在公开的中国非物质文化遗产图像数据集上区分五大类、20 个子类的西南传统工艺；论文报告准确率 96.83%。这是作者在特定整理数据集上的结果，说明图像检索与数字保存的一种技术路径，并不代表模型能在未经控制的真实场景中普遍识别文化遗产。

Researchers introduce SW-CraftNet and use a public Chinese intangible-cultural-heritage image dataset to distinguish five broad categories and 20 subcategories of traditional crafts from southwest China. The paper reports 96.83% accuracy. This author-reported result on a curated dataset suggests one route for image retrieval and digital preservation; it does not establish universal recognition in uncontrolled settings.

MOO-IDS 在 ORNL Automotive Dynamometer（ROAD）数据集上兼顾检测表现、模型体积与推理延迟。作者报告加权 F1 为 0.9907、模型大小 0.199 MB，并称延迟有所降低。结果来自已知实验数据和攻击类型，尚不能证明它在不同车辆、未知攻击或生产环境中保持同样表现。

MOO-IDS is designed to balance detection quality, model size, and inference latency on the ORNL Automotive Dynamometer (ROAD) dataset. The authors report a weighted F1 of 0.9907, a 0.199 MB model, and reduced latency. These results cover known experimental data and attack types; they do not establish equal performance across other vehicles, novel attacks, or production systems.

研究比较放射组学、深度学习与常规 PET/CT 指标预测非小细胞肺癌预后，纳入 150 名未治疗患者。模型内部 AUC 为 0.75–0.80，初始外部验证降至 0.58–0.68；在重采样分析中，仅使用总病灶糖酵解量的简单模型以 AUC 0.83 居首。论文说明复杂模型未必优于强基线，也仍只是技术可行性研究。

The study compares radiomics, deep learning, and conventional PET/CT markers for prognosis in 150 treatment-naive patients with non-small-cell lung cancer. Internal AUCs were 0.75–0.80, falling to 0.58–0.68 in the initial external test; under resampled analysis, a simple total-lesion-glycolysis model led with an AUC of 0.83. The work shows why complex models need strong baselines and remains a technical feasibility study, not a clinical tool.

研究使用机器学习分析和优化机器人冷金属过渡电弧增材制造中 AZ31 镁合金的焊道几何，把工艺参数与成形结果联系起来。论文展示的是一种面向特定材料、设备和实验条件的过程建模方法；在其他合金、机器或规模上应用前，仍需重新标定和验证。

The study uses machine learning to analyze and optimize bead geometry for AZ31 magnesium alloy in robotic cold-metal-transfer wire-arc additive manufacturing, linking process settings to deposited shape. It demonstrates process modeling for a particular material, machine, and experimental setup; other alloys, equipment, and production scales would require recalibration and validation.

研究把从头算分子动力学与机器学习分子动力学结合，用于刻画 SF₆/N₂ 混合物的扩散行为。这里的 AI 不是生成内容，而是学习近似原子间相互作用，以扩大可模拟的时间和体系尺度；结论仍受训练构型、物理假设和所研究混合物范围约束。

Researchers combine ab initio molecular dynamics with machine-learning molecular dynamics to characterize diffusion in SF₆/N₂ mixtures. Here AI does not generate content; it learns an approximation to interatomic interactions so simulations can reach larger time and system scales. The findings remain bounded by the training configurations, physical assumptions, and mixture conditions studied.

研究将递归图与同步压缩小波变换生成的脑电表示融合，再用 ResNet-18–ViT 和 EfficientNet-B0–ViT 混合架构分类，并加入解释性分析。结果属于回顾性数据集实验；脑电采集差异、样本代表性和外部验证都会影响迁移，因此不能把该模型当作现实诊断能力。

The study fuses EEG representations built from recurrence plots and synchrosqueezed wavelet transforms, then evaluates ResNet-18–ViT and EfficientNet-B0–ViT hybrids with explainability analysis. This is a retrospective dataset experiment. Differences in EEG acquisition, sample representation, and external validation affect transfer, so the model should not be treated as a real-world diagnostic capability.