ClimODE: Climate and Weather Forecasting with Physics-informed Neural ODEs
Authors: Yogesh Verma, Markus Heinonen, Vikas Garg
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our approach outperforms existing data-driven methods in global and regional forecasting with an order of magnitude smaller parameterization, establishing a new state of the art. |
| Researcher Affiliation | Collaboration | Yogesh Verma, Markus Heinonen Department of Computer Science Aalto University, Finland {yogesh.verma,markus.o.heinonen}@aalto.fi Vikas Garg Yai Yai Ltd and Aalto University vgarg@csail.mit.edu |
| Pseudocode | No | The paper does not contain any sections explicitly labeled 'Pseudocode' or 'Algorithm', nor does it present structured steps in a pseudocode format. |
| Open Source Code | Yes | Our physics-inspired model enables efficient training from scratch on a single GPU and comes with an open-source Py Torch implementation on Git Hub.1 1https://github.com/Aalto-QuML/ClimODE |
| Open Datasets | Yes | We use the preprocessed 5.625 resolution and 6 hour increment ERA5 dataset from Weather Bench (Rasp et al., 2020) in all experiments. We consider K = 5 quantities from the ERA5 dataset: ground temperature (t2m), atmospheric temperature (t), geopotential (z), and ground wind vector (u10, v10) and normalize the variables to [0, 1] via min-max scaling. |
| Dataset Splits | Yes | We use ten years of training data (2006-15), the validation data is 2016 as validation, and two years 2017-18 as testing data. |
| Hardware Specification | Yes | The whole model training and inference is conducted on a single 32GB NVIDIA V100 device. |
| Software Dependencies | No | The model is implemented in Py Torch (Paszke et al., 2019) utilizing torchdiffeq (Chen et al., 2018) to manage our data and model training. While software is mentioned, specific version numbers for PyTorch and torchdiffeq are not provided. |
| Experiment Setup | Yes | We used Cosine-Annealing-LR2 scheduler for the learning rate and also for the variance weight λσ for L2 norm shown in Fig. 7 in the loss in Eq. 12. We trained our model for 300 epochs, and the scheduler variation is shown below. |