S3GM: Learning spatiotemporal dynamics with a pretrained generative model
收藏资源简介:
Datasets of Kuramoto-Sivashinsky equation (KSE) and Kolmogorov flow. Description of KSE data: Each file for KSE datasets contains 4 dimensions in the following order: (B*V)*T*X*C. Details are listed in the following table: B number of varying initial conditions V number of varying parameters T number of temporal frames X spatial resolution C number of variables in solution (C = 1 for KSE) values of parameter used to generate training dataset 1.0, 1.2, 1.4, 1.6, 1.8, 2.0, 2.2, 2.4, 2.6, 2.8, 3.0, 3.2, 3.4, 3.6, 3.8, 4.0, 4.2, 4.4, 4.6, 4.8, 5.0 values of parameter used to generate test dataset 1.1, 2.5, 3.2 Description of Kolmogorov flow data: Each file for Kolmogorov flow contains 5 dimensions inthe following order: (B*Re*K)*T*X*X*C. Details are listed in the following table: B number of varying initial conditions Re number of varying Reynolds numbers K number of varying source terms (controled by the value of k) T number of temporal frames X spatial resolution C number of variables in solution (C = 2 for Kolmogorov flow) values of Reynolds number used to generate training dataset 100, 150, 200, 250, 300, 350, 400, 450, 500, 550, 600, 650, 700, 750, 800, 850, 900, 950, 1000, 1050 values of Reynolds number used to generate test dataset 50, 125, 575, 1100, 1500 values of k used to generate training dataset 2, 3, 4, 5, 6, 7, 8 values of k used to generate test dataset 2, 4, 6, 8 Description of ERA5 data: Training and testing dataset for ERA5 contains 5 dimensions inthe following order: 1*T*X*X*C, which is manually collected from https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=download. Note that the quantities in the datasets are already rescaled (the scale factors are saved in the scalar_era5.npy file, which is a 4x2 array recording the means and stds for the 4 quantities we used). Details are listed in the following table: T number of temporal frames X spatial resolution C number of variables in solution (C = 4 for ERA5) time span for training dataset 1979-2022 time span for test dataset 2023 Pretrained checkpoints: The .zip file contains the pretrained checkpoints for KSE, Kolmogorov flow and ERA5. Within the .zip file, the folder'kse_v0' is the checkpoint for KSE, 'kol_v0' is the checkpoint for Kolmogorov flow, and 'era5_v0' is the checkpoint for ERA5. Source code: The source code is upload as Github repository in https://github.com/lzy12301/S3GM
本数据集包含仓本-西瓦辛斯基方程(Kuramoto-Sivashinsky Equation, KSE)与柯尔莫哥洛夫流(Kolmogorov flow)两类数据。 ### KSE数据集说明 KSE数据集的每个文件均包含4个维度,维度顺序为:(B×V)×T×X×C。各维度的详细说明如下: - B:可变初始条件的数量 - V:可变参数的数量 - T:时间帧数 - X:空间分辨率 - C:解的变量数(KSE的C=1) 用于生成训练集的参数取值:1.0、1.2、1.4、1.6、1.8、2.0、2.2、2.4、2.6、2.8、3.0、3.2、3.4、3.6、3.8、4.0、4.2、4.4、4.6、4.8、5.0 用于生成测试集的参数取值:1.1、2.5、3.2 ### 柯尔莫哥洛夫流数据集说明 柯尔莫哥洛夫流数据集的每个文件包含5个维度,维度顺序为:(B×Re×K)×T×X×X×C。各维度的详细说明如下: - B:可变初始条件的数量 - Re:可变雷诺数的数量 - K:可变源项的数量(由参数k的取值控制) - T:时间帧数 - X:空间分辨率 - C:解的变量数(柯尔莫哥洛夫流的C=2) 用于生成训练集的雷诺数取值:100、150、200、250、300、350、400、450、500、550、600、650、700、750、800、850、900、950、1000、1050 用于生成测试集的雷诺数取值:50、125、575、1100、1500 用于生成训练集的k取值:2、3、4、5、6、7、8 用于生成测试集的k取值:2、4、6、8 ### ERA5数据集说明 ERA5数据集的训练集与测试集均包含5个维度,维度顺序为:1×T×X×X×C。该数据集手动采集自https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels?tab=download。需注意,数据集内的物理量已完成归一化处理,归一化因子保存在scalar_era5.npy文件中,该文件为4×2的数组,记录了所使用的4类物理量的均值与标准差。各维度的详细说明如下: - T:时间帧数 - X:空间分辨率 - C:解的变量数(ERA5的C=4) 训练集的时间跨度:1979-2022 测试集的时间跨度:2023 ### 预训练权重 本压缩包包含KSE、柯尔莫哥洛夫流与ERA5三类数据对应的预训练权重。压缩包内,文件夹kse_v0对应KSE的预训练权重,kol_v0对应柯尔莫哥洛夫流的预训练权重,era5_v0对应ERA5的预训练权重。 ### 源代码 源代码已上传至GitHub仓库:https://github.com/lzy12301/S3GM



