Sovereign 8s.is mirror of standardgalactic/physical-representation-learning
  • Python 98.6%
  • Shell 1.4%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-03-18 18:36:16 -04:00
configs initial commit 2026-03-18 18:21:19 -04:00
physics_jepa add requirements.txt 2026-03-18 18:36:16 -04:00
scripts add requirements.txt 2026-03-18 18:36:16 -04:00
.gitignore initial commit 2026-03-18 18:21:19 -04:00
README.md add requirements.txt 2026-03-18 18:36:16 -04:00
requirements.txt add requirements.txt 2026-03-18 18:36:16 -04:00

This is the official code repository for the paper Representation Learning for Spatiotemporal Physical Systems.

Installation

Requirements: Python 3.10+, PyTorch 2.0+ with CUDA.

Clone the repository and install dependencies:

git clone https://github.com/your-org/physics_jepa_public
cd physics_jepa_public
pip install torch torchvision einops omegaconf wandb tqdm h5py psutil scikit-learn timm the-well
pip install -e .

Environment setup

Edit scripts/env_setup.sh to activate your virtual environment and set the path to The Well datasets. This file is sourced automatically by all scripts. The THE_WELL_DATA_DIR variable is required by all training and finetuning scripts that use The Well data.

Training

1. JEPA pretraining

Pretrain a convolutional JEPA encoder on a physics dataset using the scripts in scripts/<dataset>/:

Dataset Script
Shear flow scripts/shear_flow/run_train_jepa.sh
Rayleigh-Bénard scripts/rayleigh_benard/run_train_jepa.sh
Active matter scripts/active_matter/run_train_jepa.sh

Config fields out_path and cache_path control where checkpoints and dataset caches are written. Key training hyperparameters (learning rate, number of epochs, noise level, etc.) are set in the train: block of the corresponding config. Config fields can be overridden from the command line by passing key=value arguments to the script, e.g.:

scripts/shear_flow/run_train_jepa.sh train.num_epochs=10 train.lr=5e-4

2. VideoMAE finetuning (baseline)

Fine-tune a pretrained VideoMAE backbone for physical parameter estimation. Set the CHECKPOINT_PATH environment variable to the pretrained VideoMAE checkpoint and run the appropriate script:

Dataset Script
Shear flow scripts/shear_flow/run_finetune_videomae.sh
Rayleigh-Bénard scripts/rayleigh_benard/run_finetune_videomae.sh
Active matter scripts/active_matter/run_finetune_videomae.sh

3. JEPA finetuning (parameter estimation)

Fine-tune a pretrained JEPA encoder for physical parameter estimation. Set CHECKPOINT_PATH to a saved encoder checkpoint and run the appropriate script:

Dataset Script
Shear flow scripts/shear_flow/run_finetune_jepa.sh
Rayleigh-Bénard scripts/rayleigh_benard/run_finetune_jepa.sh
Active matter scripts/active_matter/run_finetune_jepa.sh

The same configs used for pretraining are reused here; the ft: block controls finetuning hyperparameters. A multi-GPU variant is available at scripts/shear_flow/run_finetune_jepa_ddp.sh.

Baselines

4. DISCO finetuning

DISCO is a latent-space parameter estimation baseline. It operates on precomputed DISCO latent representations rather than raw data. Pass the path to a directory of DISCO inference outputs as the first argument:

scripts/run_finetune_disco.sh /path/to/disco_inference_shear_flow

The data directory name must match one of the dataset keys in physics_jepa/baselines/disco.py (e.g. disco_inference_shear_flow, disco_inference_rayleigh_benard, disco_inference_active_matter).

5. MPP finetuning

Fine-tune a pretrained MPP (Multiple Physics Pretraining) model for physical parameter estimation. Pass the dataset name and path to a pretrained MPP checkpoint:

scripts/run_mpp_param_estimation.sh shear_flow /path/to/MPP_AViT_Ti

--dataset_name should match the corresponding dataset directory name in THE_WELL_DATA_DIR. The checkpoint save directory can be controlled via the CHECKPOINT_DIR environment variable (defaults to ./checkpoints).