Learning Efficient Surrogate Dynamic Models with Graph Spline Networks
Authors: Chuanbo Hua, Federico Berto, Michael Poli, Stefano Massaroli, Jinkyoo Park
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate GRAPHSPLINENETS on five dynamical systems of increasing challenge. Simulated datasets consist of three PDEs (partial differential equations), including Heat, Damped Wave, and Navier-Stokes Equations. The two empirical datasets include the Ocean and Black Sea datasets; the Black Sea introduces significant challenges in scalability in the number of nodes, complex dynamics, irregular boundary, and non-uniform meshing. |
| Researcher Affiliation | Academia | Chuanbo Hua KAIST Diffeq ML, AI4CO Federico Berto KAIST Diffeq ML, AI4CO Michael Poli Stanford University Diffeq ML Stefano Massaroli Mila and University of Montreal Diffeq ML Jinkyoo Park KAIST OMELET |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | We value open reproducibility and make our code publicly available5. |
| Open Datasets | Yes | Simulated datasets consist of three PDEs (partial differential equations), including Heat, Damped Wave, and Navier-Stokes Equations. The two empirical datasets include the Ocean and Black Sea datasets; the Black Sea introduces significant challenges in scalability in the number of nodes, complex dynamics, irregular boundary, and non-uniform meshing. ... The ocean currents dataset (Nardelli et al., 2013; Marullo et al., 2014)... The Black Sea dataset is composed of daily real-world measurements of ocean currents and temperatures (Ciliberti et al., 2021). ... The dataset is available to download from the Copernicus Institute9. |
| Dataset Splits | Yes | For all datasets, we used the split of 5 : 1 : 1 for training, validating, and testing for a fair comparison. |
| Hardware Specification | Yes | Hardware Experiments were carried out on a machine equipped with an INTEL CORE I9 7900X CPU with 20 threads and a NVIDIA RTX A5000 graphic card with 24 GB of VRAM. |
| Software Dependencies | No | Software-wise, we used FEni CS (Logg et al., 2012) for Finite Element simulations for the heat equation experiments and Py Torch (Paszke et al., 2019) for simulating the damped wave and Navier-Stokes equations. We employed the Deep Graph Library (DGL) (Wang et al., 2020) for graph neural networks. ... combines the Py Torch Lightning library (Falcon et al., 2019) for efficiency with modular Hydra configurations (Yadan, 2019). |
| Experiment Setup | Yes | A batch size of 32 was used for all experiments, and the models were trained for up to 5000 epochs with early stopping. We used the Adam optimizer (Kingma and Ba, 2014) with an initial learning rate of 0.001 and a step scheduler with a 0.85 decay rate every 500 epoch. |