Graph Network Simulator Datasets

The synthesis of graph networks and deep learning models present a unique opportunity to scale computationally intensive simulations beyond current capabilities. In their original work, Sanchez-Gonzalez et al. (2020) demonstrated that using graph networks with relatively simple deep learning models, so called “Graph Network-based Simulators” (GNS), could be used to model complex physics involving interactions between fluids, rigid solids, and deformable materials. Alongside their paper, Sanchez-Gonzalez et al. (2020) released a proof-of-concept implementation of their GNS written in TensorFlow version 1 and several datasets in the .tfrecord format for model training, validation, and testing. The data in this project includes three of the datasets from Sanchez-Gonzalez et al. (2020), in particular, Sand, SandRamps, and WaterDropSample, converted to the .npz format and designed to integrate with a recently released port of the GNS from TensorFlow version 1 to PyTorch (Kumar and Vantassel, 2022). The PyTorch version of the GNS code is available at https://github.com/geoelements/gns.

图网络（graph networks）与深度学习模型的融合，为突破现有算力限制、扩展计算密集型模拟的规模提供了独特机遇。在其开创性研究中，Sanchez-Gonzalez等人（2020）证实，采用搭配相对简单深度学习模型的图网络——即所谓的「基于图网络的模拟器（Graph Network-based Simulators，GNS）」——可用于建模包含流体、刚性固体与可变形材料间相互作用的复杂物理系统。伴随该论文发布的，还有Sanchez-Gonzalez等人（2020）基于TensorFlow 1版本编写的GNS概念验证实现，以及若干用于模型训练、验证与测试的.tfrecord格式数据集。本项目包含Sanchez-Gonzalez等人（2020）发布的三类数据集，具体为Sand、SandRamps与WaterDropSample，已转换为.npz格式，并适配了近期将GNS从TensorFlow 1版本移植至PyTorch的开源项目（Kumar与Vantassel，2022）。该PyTorch版本的GNS代码可通过以下链接获取：https://github.com/geoelements/gns。

搜集汇总

背景与挑战

背景概述

该数据集包含三个物理模拟训练数据集(Sand, SandRamps, WaterDropSample)，这些数据从TensorFlow的.tfrecord格式转换为.npz格式，用于支持基于PyTorch框架的图网络模拟器(GNS)开发。数据集源自Sanchez-Gonzalez等人关于图网络与深度学习结合的研究工作，旨在扩展计算密集型物理模拟的能力。