PraLak/Hybrid_Neural_World_Models

该数据集名为混合神经世界模型，包含了三个物理系统的训练、验证、测试和分布外轨迹数据，用于支持神经代理模型的研究。具体包括：1) 反应扩散PDE（Belousov-Zhabotinsky）系统，基于Oregonator模型，在256x256周期性网格上模拟，每个轨迹包含201个时间帧，状态形状为(2, 256, 256)的浮点数，表示两种化学浓度；2) 可压缩气体流动PDE（Euler 2D）系统，使用MUSCL-Hancock + HLLC方案在128x128网格上求解，每个轨迹包含100个时间帧，状态为四个保守场(ρ, ρvx, ρvy, E)，展平为每时间步16384个值；3) 刚体碰撞ODE（Ball 3D）系统，基于MuJoCo模拟，每个轨迹包含101个时间帧，状态为9维向量，包括位置、线速度和角速度。数据集还提供了分布外划分（OOD-near和OOD-far），用于测试模型的鲁棒性。数据分割大小明确，训练集包含1200（Oregonator）、800（Euler 2D）和1000（Ball 3D）个独立轨迹。数据集旨在用于基准测试神经代理模型、研究无标签不确定性方法以及跨系统泛化研究。The dataset is named Hybrid Neural World Models and contains training, validation, test, and out-of-distribution trajectories for three physical systems: 1) Reaction-diffusion PDE (Belousov-Zhabotinsky) based on the Oregonator model, simulated on a 256x256 periodic grid with 201 saved frames per trajectory, state shape (2, 256, 256) float32 representing two chemical concentrations; 2) Compressible gas flow PDE (Euler 2D) solved with an MUSCL-Hancock + HLLC scheme on a 128x128 grid, each trajectory with 100 saved frames of four conservative fields (ρ, ρvx, ρvy, E) flattened to 16,384 per timestep; 3) Rigid-body bouncing ODE (Ball 3D) simulated in MuJoCo, each trajectory with 101 frames of a 9-vector state including position, linear velocity, and angular velocity. The dataset includes OOD splits (near and far) for robustness testing. Split sizes are specified: train set has 1,200 (Oregonator), 800 (Euler 2D), and 1,000 (Ball 3D) independent trajectories. It is intended for benchmarking neural surrogates, studying label-free uncertainty methods, and cross-system generalization research.