Research Experiences for Undergraduates (REU), NSF NHERI 2024: Hyperparameter Optimization of Graph Neural Network Simulator for Liquid and Granular Flow

This project aims to optimize the Hyperparameters of the Graph Neural Network Simulator (GNS) to address the limitations of GNS -simulation scalability and error accumulation from longer rollouts- which restricts its application in engineering fields. GNS was developed to accurately simulate granular and liquid flow while maintaining a low computational cost and generalizability. It is used by engineers to mitigate the effects of natural hazards such as landslides. This project data can be used to further improve GNS model and increase its application in mitigating natural hazards. It would be the first time GNS's hyperparameters were optimized, which is an important process in improving machine learning models. The audience would be natural hazards engineers.

本项目旨在对图神经网络模拟器（Graph Neural Network Simulator, GNS）的超参数（Hyperparameters）进行优化，以解决其模拟可扩展性不足、长序列推演时误差累积等局限性——上述问题限制了该模型在工程领域的应用。GNS的研发目标是在保持较低计算成本与泛化能力的前提下，精准模拟颗粒流与液流的运动过程。工程人员可借助该模型缓解滑坡等自然灾害所造成的危害。本项目生成的数据集可用于进一步优化GNS模型，拓展其在自然灾害减灾领域的应用空间。本次针对GNS的超参数优化尚属首次，而超参数调优是提升机器学习模型性能的关键流程。本项目的目标受众为自然灾害防控工程师。