Physics-Embedded Neural Networks: Graph Neural PDE Solvers with Mixed Boundary Conditions

Authors: Masanobu Horie, NAOTO MITSUME

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using numerical experiments, we demonstrate the proposed model s validity, expressibility, and computational efficiency. We use two types of datasets: 1) the gradient dataset to verify the correctness of NIso GCN and 2) the incompressible flow dataset to demonstrate the speed and accuracy of the model.
Researcher Affiliation Collaboration Masanobu Horie RICOS Co. Ltd. University of Tsukuba horie@ricos.co.jp Naoto Mitsume University of Tsukuba mitsume@kz.tsukuba.ac.jp
Pseudocode No No structured pseudocode or algorithm blocks were found.
Open Source Code Yes The code is available at https://github.com/yellowshippo/penn-neurips2022.
Open Datasets Yes Our training, validation, and test datasets consisted of 100 samples. and Training, validation, and test datasets consisted of 203, 25, and 25 samples, respectively.
Dataset Splits Yes Our training, validation, and test datasets consisted of 100 samples. and Training, validation, and test datasets consisted of 203, 25, and 25 samples, respectively.
Hardware Specification Yes All models were trained for up to 24 hours using one GPU (NVIDIA A100 for NVLink 40Gi B HBM2). and All computation was done using one core of Intel Xeon CPU E5-2695 v2@2.40GHz.
Software Dependencies No The paper mentions 'Open FOAM' for dataset generation and states that 'The implementation of our model is based on the original Iso GCN s code,' but does not provide specific version numbers for software dependencies like Python, PyTorch, or other libraries used in the implementation.
Experiment Setup Yes We encoded each feature in a 4, 8, or 16-dimensional space. and We looped the solver for pressure five times and four or eight times for velocity. and We tested various time window sizes such as 2, 4, 10, and 20... The numbers of hidden features, 32, 64, and 128, were tested. All models were trained for up to 24 hours...