LE-PDE: Learning to Accelerate Forward Simulation and Inverse Optimization of PDEs via Latent Global Evolution
Paper | Poster | Slide | Project Page
Official repo for the paper Learning to Accelerate Partial Differential Equations via Latent Global Evolution
Tailin Wu, Takashi Maruyama, Jure Leskovec
NeurIPS 2022
It introduces a simple, fast and scalable LE-PDE method to accelerate the simulation and inverse optimization of PDEs, which are crucial in many scientific and engineering applications (e.g., weather forecasting, material science, engine design).
LE-PDE achieves up to 128x reduction in the dimensions to update, and up to 15x improvement in speed, while achieving competitive accuracy compared to state-of-the-art deep learning-based surrogate models (e.g., FNO, MP-PDE).
- First clone the directory. Then run the following command to initialize the submodules:
git submodule init; git submodule update
(If showing error of no permission, need to first add a new SSH key to your GitHub account.)
- Install dependencies.
First, create a new environment using conda (with python >= 3.7). Then install pytorch, torch-geometric and other dependencies as follows (the repository is run with the following dependencies. Other version of torch-geometric or deepsnap may work but there is no guarentee.)
Install pytorch (replace "cu113" with appropriate cuda version. For example, cuda11.1 will use "cu111"):
pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2+cu113 -f https://download.pytorch.org/whl/torch_stable.html
Install torch-geometric. Run the following command:
pip install torch-scatter==2.0.9 -f https://data.pyg.org/whl/torch-1.10.2+cu113.html
pip install torch-sparse==0.6.12 -f https://data.pyg.org/whl/torch-1.10.2+cu113.html
pip install torch-geometric==1.7.2
pip install torch-cluster==1.5.9 -f https://data.pyg.org/whl/torch-1.10.2+cu113.html
Install other dependencies:
pip install -r requirements.txt
If you want to do inverse optimization, run also the following command inside the root directory of PDE_Control repository:
pip install PDE-Control/PhiFlow/[gui] jupyterlab
The dataset files can be downloaded via this link.
- To run 1D experiment, download the files under "mppde1d_data/" in the link into the "data/mppde1d_data/" folder in the local repo.
- To run 2D experiment, download the files under "fno_data/" in the link into the "data/fno_data/" folder in the local repo.
- To run inverse optimization experiment, download the files under "movinggas_data/" in the link into the "data/movinggas_data/" folder in the local repo.
Below we provide example commands for training LE-PDEs. For all the commands that reproduce the experiments in the paper, see the results/README.md.
An example 1D training command is:
python train.py --exp_id=le-pde-1d --date_time=2022-11-21 --dataset=mppde1d-E2-50 --n_train=-1 --time_interval=1 --save_interval=2 --algo=contrast --no_latent_evo=False --encoder_type=cnn-s --input_steps=1 --evolution_type=mlp-3-elu-2 --decoder_type=cnn-tr --encoder_n_linear_layers=0 --n_conv_blocks=4 --n_latent_levs=1 --n_conv_layers_latent=3 --channel_mode=exp-32 --is_latent_flatten=True --evo_groups=1 --recons_coef=1 --consistency_coef=1 --contrastive_rel_coef=0 --hinge=0 --density_coef=0.001 --latent_noise_amp=1e-5 --normalization_type=gn --latent_size=128 --kernel_size=4 --stride=2 --padding=1 --padding_mode=zeros --act_name=elu --multi_step=1^2^3^4 --latent_multi_step=1^2^3^4 --temporal_bundle_steps=25 --use_grads=False --use_pos=False --is_y_diff=False --loss_type=rmse --loss_type_consistency=mse --batch_size=16 --val_batch_size=16 --epochs=50 --opt=adam --weight_decay=0 --seed=0 --gpuid=9 --id=0 --verbose=1 --save_iterations=1000 --latent_loss_normalize_mode=targetindi --n_workers=0 --static_encoder_type=param-0 --static_latent_size=3 --gpuid=0
An example 2D training command is:
python train.py --exp_id=le-pde-2d --date_time=2022-11-21 --dataset=fno-4 --n_train=-1 --time_interval=1 --save_interval=10 --algo=contrast --no_latent_evo=False --encoder_type=cnn-s --input_steps=10 --evolution_type=mlp-3-elu-2 --decoder_type=cnn-tr --encoder_n_linear_layers=0 --n_conv_blocks=4 --n_latent_levs=1 --n_conv_layers_latent=3 --channel_mode=exp-16 --is_latent_flatten=True --evo_groups=1 --recons_coef=1 --consistency_coef=1 --contrastive_rel_coef=0 --hinge=0 --density_coef=0.001 --latent_noise_amp=1e-5 --normalization_type=gn --latent_size=256 --kernel_size=4 --stride=2 --padding=1 --padding_mode=zeros --act_name=elu --multi_step=1^2:0.1^3:0.1^4:0.1 --latent_multi_step=1^2^3^4 --use_grads=False --use_pos=False --is_y_diff=False --loss_type=mse --loss_type_consistency=mse --batch_size=20 --val_batch_size=20 --epochs=200 --opt=adam --weight_decay=0 --seed=0 --gpuid=9 --id=0 --verbose=1 --save_iterations=400 --latent_loss_normalize_mode=targetindi --n_workers=0 --gpuid=0
An example command of training for smoke simulation (used for inverse design) is:
python train.py --exp_id=le-pde-smoke --date_time=2022-11-21 --dataset=movinggas --n_train=-1 --time_interval=1 --save_interval=10 --algo=contrast --reg_type=None --reg_coef=0 --is_reg_anneal=True --no_latent_evo=False --encoder_type=cnn-s --evolution_type=mlp-3-elu-2 --decoder_type=cnn-tr --encoder_n_linear_layers=0 --n_conv_blocks=4 --n_latent_levs=1 --n_conv_layers_latent=3 --channel_mode=exp-16 --is_latent_flatten=True --evo_groups=1 --recons_coef=1 --consistency_coef=1 --contrastive_rel_coef=0 --hinge=0 --density_coef=0.001 --latent_noise_amp=1e-5 --normalization_type=gn --latent_size=128 --kernel_size=4 --stride=2 --padding=1 --padding_mode=zeros --act_name=elu --multi_step=1^2:0.1^3:0.1^4:0.1 --latent_multi_step=1^2^3^4 --use_grads=False --use_pos=False --is_y_diff=False --loss_type=mse --loss_type_consistency=mse --batch_size=16 --val_batch_size=16 --epochs=100 --opt=adam --weight_decay=0 --seed=0 --id=0 --verbose=1 --save_iterations=1000 --latent_loss_normalize_mode=targetindi --n_workers=0 --static_encoder_type="cnn-s" --static_latent_size=16 --gpuid=0
The results are saved under results/{--exp_id}_{--date_time}/ (here --exp_id and --date_time are according to the command for training). Each experiment file has the following suffix: "*Hash_{hash}_{machine-name}.p". The hash (e.g., "Un6ae7ja"), is uniquely generated according to all the configurations of the argparse (if any argument is different, it will result in a different hash).
inverse_design.ipynb is a script file for inverse design to optimize the boundary condition. exp_id and data_time need to be provided to identify folder storing a model with which you perform inverse design. They should be part of the folder's name as described above.
To analyze the results, use the following notebooks:
- 1D: analysis_1d.ipynb
- 2D: analysis_2d.ipynb
- 3D: analysis_3d.ipynb
- Inverse optimization: analysis_inverse.ipynb
Pre-trained experiment files can also be downloaded here (put it under result/, and also change the dirname in the analysis notebook accordingly).
- LAMP (ICLR 2023 spotlight): first fully DL-based surrogate model that jointly optimizes spatial resolutions to reduce computational cost and learns the evolution model, learned via reinforcement learning.
If you find our work and/or our code useful, please cite us via:
@inproceedings{wu2022learning,
title={Learning to accelerate partial differential equations via latent global evolution},
author={Wu, Tailin and Maruyama, Takashi and Leskovec, Jure},
booktitle={Neural Information Processing Systems},
year={2022},
}