MIT's WorldParticle simulates multiple materials with one AI model
MIT developed WorldParticle, an AI system that represents objects as 3D particles and predicts their next physical states. According to TechNews, the model can simulate six types of physical behavior, including granular materials, liquids, deformable solids and changing protein structures, with protein simulation reportedly about 100 times faster than traditional methods.
WorldParticle could unify simulations that currently require separate material-specific tools, offering a step toward foundation models for physics.

- )
麻省理工學院(MIT)開發新 AI 系統「WorldParticle」,把物體拆解成大量 3D 粒子並預測下個運動狀態,可用同一模型模擬沙土、液體、可變形固體與蛋白質等多種材料的物理行為。
AI 影片產生器最能看出 AI 仍難預測並正確呈現現實物體的物理行為。Notebookcheck 作者測試 RefVid(Kling 4 模型)時,人會卡進桌子裡、筆電變成三摺裝置。準確模擬物理世界與各式各樣物體至今仍是難題,不同用途往往得靠各自專門打造的模擬軟體。
WorldParticle 方法不同:不為每種材料分別寫物理方程式與客製化的模擬,而是把各種物體拆成大量微小 3D 粒子。每顆粒子都有空間位置、速度、物理性質,以及與其他粒子或外力的交互作用。就像 ChatGPT 預測下個字,WorldParticle 預測的是粒子的下個運動狀態。