文章

发布于

日本MMX探测器如何在没有重力的火卫一上着陆

日本MMX探测器于10月20日搭乘H3火箭发射，计划2027年抵达火星，对火卫一进行多次着陆并取样返回。火卫一重力仅为地球的1/1800，表面是松散碎石堆，MMX将先释放23公斤的Idefix巡视器，以低于1米每秒的速度弹跳着陆，再用机械臂取芯和氮气气动两种方式采集至少10克样品，预计2031年下半年送回地球。

An photo shows a lumpy dark spot in front of a glowing orb.
A photo shows an irregular gray and brown surface with a large cavity on one side.
A tall spacecraft wrapped in gold and silver thermal insulation stands upright in a clean room.
A photo shows two spacecraft assemblies inside a clean room, with two people in white suits standing nearby.
A rocket rises vertically into the dark sky above a launchpad.
A schematic shows Mars at the center with curved spacecraft trajectories looping around it, three labeled maneuver points, and the orbits of Phobos and Deimos.
Two spacecraft instruments are shown side by side: a cylindrical metallic camera assembly on the left and a box-shaped, gold-covered laser instrument on the right.
A box-shaped rover wrapped in gold and black materials has folded wheel assemblies and protruding structural components.
A small, dark rover with folded wheels is attached beneath a larger spacecraft structure covered in gold and silver insulation.
A four-wheeled rover model stands on simulated rocky terrain, with four rectangular solar panels spread out horizontally above its body.
A compact rover is mounted beneath a tall robotic sampling assembly, with metallic structural supports and gold thermal insulation visible around both.
An artist's rendering shows a spacecraft resting on a gray, dusty surface.
A rounded, dome-shaped metallic capsule with a reflective gold surface sits inside a support frame.

来源：IEEE Spectrum · 阅读原文 · 前沿热点

发布于

MIT 团队提出气候模拟器训练新方法，用少量合成数据提升预测可靠性

MIT 研究人员开发了一种训练技术，用精心挑选的少量合成数据训练气候模拟器，使其能准确预测训练中未见过的情景，而传统模拟器无法适应新情况。该算法只需运行一次完整地球系统模型即可生成训练数据，而非通常的六到八次，从而大幅缩短模拟器制作时间。相关论文本周发表在 PNAS，研究由 Schmidt Sciences 通过 MIT Grand Challenges 的 BC3 项目支持。

来源：MIT 研究新闻 · 阅读原文 · 前沿热点

发布于

Hugging Face Transformers v5.22.0 支持 Mistral Large 4 万亿参数多模态模型

Hugging Face Transformers 发布 v5.22.0，新增对 Mistral Large 4（Le Chonk）的支持。该模型总参数约 1T、每 token 激活约 52B，采用 128 个路由专家、每 token 激活 4 个加 1 个共享专家的稀疏 MoE 结构，支持 1M 上下文和文本加图像输入，推理强度分 low、medium、high 三档。

来源：Transformers 正式版本发布 · 阅读原文 · 前沿热点