Skip to content

[ICLR23] First deep learning-based surrogate model that jointly learns the evolution model and optimizes computational cost via remeshing

License

Notifications You must be signed in to change notification settings

snap-stanford/lamp

Repository files navigation

LAMP: Learning Controllable Adaptive Simulation for Multi-resolution Physics (ICLR 2023 Notable-Top-25%)

Paper | arXiv | Poster | Slides | Tweet | Project Page

Official repo for the paper Learning Controllable Adaptive Simulation for Multi-resolution Physics
Tailin Wu*, Takashi Maruyama*, Qingqing Zhao*, Gordon Wetzstein, Jure Leskovec
ICLR 2023 Notable-Top-25%.

It is the first fully DL-based surrogate model that jointly learns the evolution model, and optimizes spatial resolutions to reduce computational cost, learned via reinforcement learning.

We demonstrate that LAMP is able to adaptively trade-off computation to improve long-term prediction error, by performing spatial refinement and coarsening of the mesh. LAMP outperforms state-of-the-art (SOTA) deep learning surrogate models, with an average of 33.7% error reduction for 1D nonlinear PDEs, and outperforms SOTA MeshGraphNets + Adaptive Mesh Refinement in 2D mesh-based simulations.

Learned remeshing & evolution by LAMP:

Installation

  1. 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.)

  1. 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 wanting to use wandb (--wandb=True), need to set up wandb, following this link.

If wanting to run 2d mesh-based simulation, FEniCS needs to be installed:

conda install -c conda-forge fenics

Dataset

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 mesh-based experiment, download the files under "arcsimmesh_data/" in the link into the "data/arcsimmesh_data/" folder in the local repo. Script for data generation is also provided ("datasets/datagen_square.ipynb".) To run the script, compile ARCSim v0.2.1 and place the script in ARCSim folder. The detailed explanation for the attributes for the mesh-based dataset are provided under datasets/README.md.

Training

Below we provide example commands for training LAMP.

1D nonlinear PDE:

First, pre-train the evolution model for 1D:

python train.py --exp_id=evo-1d --date_time=2023-01-01 --dataset=mppde1df-E2-100-nt-250-nx-200 --time_interval=1 --data_dropout=node:0-0.3:0.1 --latent_size=64 --n_train=-1 --save_interval=5 --test_interval=5 --algo=gnnremesher --rl_coefs=None --input_steps=1 --act_name=silu --multi_step=1^2:0.1^3:0.1^4:0.1 --temporal_bundle_steps=25 --use_grads=False --is_y_diff=False --loss_type=mse --batch_size=16 --val_batch_size=16 --epochs=50 --opt=adam --weight_decay=0 --seed=0 --id=0 --verbose=1 --n_workers=0 --gpuid=0

The learned model will be saved under ./results/{--exp_id}_{--date_time}/, where the {--exp_id} and {--date_time} are specified in the above command. The filename has the format of *{hash}_{machine_name}.p, e.g. "mppde1df-E2-100-nt-250-nx-200_train_-1_algo_gnnremesher_..._Hash_mhkVkAaz_ampere3.p", then the {hash} is mhkVkAaz and {machine_name} is ampere3, where the {hash} is uniquely determined by all the argument settings in the argparser.py (therefore, as long as any argument setting is different, the filename will be different and will not overwrite each other).

Then, jointly train the remeshing model via reinforcement learning (RL) and the evolution model. The --load_dirname below should use folder name {exp_id}_{date_time} where the evolution model is located (as specified above), and the --load_filename should use part of the filename that can uniquely identify this model file, and should include the {hash} of this model.

python train.py --load_dirname=evo-1d_2023-01-01 --load_filename=Q66bz42y --exp_id=rl-1d --date_time=2023-01-02 --wandb_project_name=rl-1d_2023-01-02 --wandb=True --dataset=mppde1df-E2-100-nt-250-nx-200 --time_interval=1 --data_dropout=None --latent_size=64 --n_train=-1 --input_steps=1 --act_name=elu --multi_step=1^2:0.1^3:0.1^4:0.1 --temporal_bundle_steps=25 --use_grads=False --is_y_diff=False --loss_type=mse --batch_size=128 --val_batch_size=128 --epochs=30 --opt=adam --weight_decay=0 --seed=0 --verbose=1 --n_workers=0 --gpuid=7 --algo=srlgnnremesher --reward_mode=lossdiff+statediff --reward_beta=0-0.5 --rl_data_dropout=uniform:2 --min_edge_size=0.0014 --rl_horizon=4 --reward_loss_coef=5 --rl_eta=1e-2 --actor_lr=5e-4 --value_lr=1e-4 --value_num_pool=1 --value_pooling_type=global_mean_pool --value_latent_size=32 --value_batch_norm=False --actor_batch_norm=True --rescale=10 --edge_attr=True --rl_gamma=0.9 --value_loss_coef=0.5 --max_grad_norm=2 --is_single_action=False --value_target_mode=vanilla --wandb_step_plot=100 --wandb_step=20 --save_iteration=1000 --save_interval=1 --test_interval=1 --gpuid=3 --lr=1e-4 --actor_critic_step=200 --evolution_steps=200 --reward_condition=True --max_action=20 --rl_is_finetune_evolution=True --rl_finetune_evalution_mode=policy:fine  --id=0

2D mesh-based simulation:

Pre-train the evolution model for 2D (need to have FEniCS installed, see "Installation" section:

export OMP_NUM_THREADS=6; python train.py --exp_id=evo-2d  --date_time=2023-01-01 --dataset=arcsimmesh_square_annotated --time_interval=2 --data_dropout=None --n_train=-1 --save_interval=5 --algo=gnnremesher-evolution --rl_coefs=None --input_steps=2 --act_name=silu --multi_step=1 --temporal_bundle_steps=1 --edge_attr=True --use_grads=False --is_y_diff=False --loss_type=l2 --batch_size=10 --val_batch_size=10 --latent_size=56 --n_layers=8 --noise_amp=1e-2 --correction_rate=0.9 --epochs=100 --opt=adam --weight_decay=0  --is_mesh=True --seed=0 --id=0 --verbose=2 --test_interval=2 --n_workers=20 --gpuid=0

Then, jointly train the remeshing model via RL and the evolution model:

export OMP_NUM_THREADS=6; python train.py --exp_id=2d_rl --wandb_project_name=2d_rerun --wandb=True --date_time=2023-02-26 --dataset=arcsimmesh_square_annotated_coarse_minlen008_interp_500 --time_interval=2 --n_train=-1 --latent_size=64 --load_dirname=evo-2d_2023_02_18 --load_filename=9UQLIKKc_ampere1 --input_steps=2 --act_name=elu --temporal_bundle_steps=1 --use_grads=False --is_y_diff=True --loss_type=l2 --epochs=300 --opt=adam --weight_decay=0 --verbose=1 --algo=srlgnnremesher --reward_mode=lossdiff+statediff --rl_data_dropout=None --min_edge_size=0.04 --actor_lr=5e-4 --value_lr=1e-4 --value_num_pool=1 --value_pooling_type=global_mean_pool --value_latent_size=64 --value_batch_norm=False --actor_batch_norm=True --rescale=10 --edge_attr=True --rl_gamma=0.9 --value_loss_coef=0.5 --max_grad_norm=20 --is_single_action=False --value_target_mode=vanilla --wandb_step_plot=50 --wandb_step=2 --id=0 --save_iteration=500 --save_interval=1 --test_interval=1 --is_mesh=True --is_unittest=False --rl_horizon=6 --multi_step=6 --rl_eta=2e-2 --reward_beta=0 --reward_condition=True --max_action=20 --rl_is_finetune_evolution=True --lr=1e-4 --actor_critic_step=200 --evolution_steps=100 --rl_finetune_evalution_mode=policy:fine --wandb=True --batch_size=64 --val_batch_size=64 --n_workers=6 --reward_loss_coef=1000 --evl_stop_gradient=True --noise_amp=0.01 --gpuid=5 --is_eval_sample=True --seed=256 --n_train=:-1 --soft_update=False --fine_tune_gt_input=True --policy_input_feature=coords --skip_coarse=False  --skip_flip=True  --processor_aggr=mean  --fix_alt_evolution_model=True

For commands for baseline models in 1D, see the README in ./MP_Neural_PDE_Solvers/.

We also provide pre-trained evolution models directly for RL training here. Put the folders in the Google doc (e.g., "evo-1d_2023-01-01" for pre-trained evolution model for 1d, "evo-2d_2023-02-18" for pre-trained evolution model for 2d) under the ./results/ folder, and can then use the RL commands above to perform joint training.

Analysis

Visualization:

Example visualization of learned remeshing & evolution:

LAMP:

LAMP (no remeshing):

MeshGraphNets + ground-truth remeshing:

MeshGraphNets + heuristic remeshing:

Ground-truth (fine-grained):

Related Projects

  • LE-PDE (NeurIPS 2022): Accelerate the simulation and inverse optimization of PDEs. Compared to state-of-the-art deep learning-based surrogate models (e.g., FNO, MP-PDE), it is up to 15x improvement in speed, while achieving competitive accuracy.

  • CinDM (ICLR 2024 spotlight): We introduce a method that uses compositional generative models to design boundaries and initial states significantly more complex than the ones seen in training for physical simulations.

  • BENO (ICLR 2024): We introduce a boundary-embedded neural operator that incorporates complex boundary shape and inhomogeneous boundary values into the solving of Elliptic PDEs.

Citation

If you find our work and/or our code useful, please cite us via:

@inproceedings{wu2023learning,
  title={Learning Controllable Adaptive Simulation for Multi-resolution Physics},
  author={Tailin Wu and Takashi Maruyama and Qingqing Zhao and Gordon Wetzstein and Jure Leskovec},
  booktitle={The Eleventh International Conference on Learning Representations},
  year={2023},
  url={https://openreview.net/forum?id=PbfgkZ2HdbE}
}

About

[ICLR23] First deep learning-based surrogate model that jointly learns the evolution model and optimizes computational cost via remeshing

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published