Training, validation, testing data, and trained model

The dataset includes the training/validation/testing data for the graph neural network (GNN)-based simulator (GNS) used in Choi and Kumar (2023) "Graph Neural Network-based Surrogate Model for Granular Flows" cited in the Related Work. Graph neural network (GNN)-based simulator (GNS) is a generalizable, efficient, and accurate machine learning (ML)-based surrogate simulator that uses Graph Neural Networks (GNNs) originally introduced in the paper "Learning to Simulate Complex Physics with Graph Networks" by DeepMind (see the Related work). GNS is a viable surrogate for numerical methods such as Material Point Method, Smooth Particle Hydrodynamics and Computational Fluid dynamics and can be extended to simulate natural hazards. GNS can handle complex boundary conditions and multi-material interactions. We improved GNS to be able to exploit distributed data parallelism to achieve fast multi-GPU training. This improved version of GNS is published in GitHub "Graph Network Simulator (GNS) and MeshNet" (see the Related Work). We investigate the performance of GNS in learning to simulate complex granular flow problems: "BarrierInteraction", "ColumnCollapseSimple", and "ColumnCollapseFrictional". The details of these granular flow problems are explained in Choi and Kumar (2023). The GNS is trained on the `train.npz`, and validated and tested on the `validation.npz` and `test.npz` uploaded in this dataset. Detailed instructions on how to train and test the GNS are explained in `README.md` file in the GitHub repository https://github.com/geoelements/gns in the Related Work. The trained model is also included in this dataset. The result shows that GNS can accurately simulate the granular flow dynamics faster than high-fidelity simulators, and is generalizable to other configurations not seen during the training. More information about the result can be found in Choi and Kumar (2023).

本数据集包含相关工作中引用的Choi与Kumar（2023）发表的《基于图神经网络的颗粒流动代理模型》一文所用的、基于图神经网络（Graph Neural Network, GNN）的模拟器（GNS）的训练、验证与测试数据。基于图神经网络（GNN）的模拟器（GNS）是一种通用、高效且精准的基于机器学习（Machine Learning, ML）的代理模拟器，其核心的图神经网络架构源自DeepMind发表的论文《用图网络学习模拟复杂物理过程》（详见相关工作部分）。GNS可作为物质点法（Material Point Method）、光滑粒子流体动力学（Smooth Particle Hydrodynamics）与计算流体动力学（Computational Fluid Dynamics）等数值方法的可行代理，还可扩展用于自然灾害模拟。它能够处理复杂边界条件与多材料交互场景。我们对GNS进行了改进，使其支持分布式数据并行训练，以实现快速多GPU训练。该改进版GNS已发布于GitHub仓库"Graph Network Simulator (GNS) and MeshNet"（详见相关工作部分）。我们探究了GNS在学习模拟复杂颗粒流动问题时的性能，所涉及的问题包括`BarrierInteraction`、`ColumnCollapseSimple`以及`ColumnCollapseFrictional`。这些颗粒流动问题的细节可参见Choi与Kumar（2023）的研究。本数据集上传了`train.npz`用于GNS的训练，以及`validation.npz`和`test.npz`用于模型的验证与测试。关于GNS的训练与测试的详细操作说明，可参见相关工作中GitHub仓库https://github.com/geoelements/gns内的`README.md`文件。本数据集还附带了训练完成的模型。实验结果表明，GNS能够以高于高保真模拟器的速度精准模拟颗粒流动动力学过程，且可泛化至训练阶段未见过的其他配置场景。更多关于实验结果的细节可参阅Choi与Kumar（2023）的研究。