EVENT DOSSIER
On September 7, 2026, arXiv cs.AI released GyroSwin, the first scalable 5D neural agent model designed to simulate nonlinear gyrokinetic plasma turbulence. This model extends the hierarchical visual Transformer to 5D, introduces cross-attention and integrated modules to handle interactions between electrostatic potential fields and distribution functions, and implements physics-inspired channel mode separation. Experimental results show that GyroSwin outperforms existing simplified numerical methods in heat flux prediction, can capture turbulence energy cascades, reduces the computational cost of full-resolution gyrokinetic calculations by three orders of magnitude while maintaining physical verifiability. The model has been tested up to a billion parameters scale, providing a scalable neural agent solution for plasma turbulence simulation.
GyroSwin 是首个可扩展的 5D 神经代理模型，用于模拟非线性陀螺动理论等离子体湍流。该模型将分层视觉 Transformer 扩展至 5D，引入交叉注意力与集成模块处理静电势场与分布函数间的 3D 至 5D 交互，并实施受非线性物理启发的通道模式分离。实验表明，GyroSwin 在热通量预测上优于广泛使用的简化数值方法，能捕捉湍流能量级联，将全分辨率非线性陀螺动理论计算成本降低三个数量级，且保持物理可验证性。该模型展示了具有前景的扩展规律，已测试至十亿参数规模，为等离子体湍流的陀螺动理论模拟提供了可扩展的神经代理方案。