This example demonstrates how to leverage the optimized model implementations in Modulus for different domains. This example showcases how you can leverage the MeshGraphNet model in Modulus for developing a DL model for predicting forces/potential for a Lennard Jones System as described in the paper here.
The goal is to train an AI model that can predict the forces on atoms of a Lennard Jones system (liquid Argon) given the positions of its atoms.
The model is trained on data generated using OpenMM MD simulator. The dataset consists of 10000 samples of the 258 atom system. For original dataset please refer the original publication and Git repo of the origial work.
The model uses a MeshGraphNet model for the prediction of forces. Since all the atoms in this system are of same type (i.e. Argon), the node encoder is dropped. The graph edges are generated based on nearest-neighbor search.
Fig. 7 Results from Modulus training for the LJ system.
To download the data, run
pip install gdown
python download_data.py
To train the model, run
python lennard_jones_system.py
Distributed Data Parallel training is enabled for this example. To run the example on multiple GPUs, run
mpirun -np <num_GPUs> python lennard_jones_system.py
If running in a docker container, you may need to include the
--allow-run-as-root in the multi-GPU run command.
While the current example trains a light-weight model that can be run on any GPU, on 8 A100s, the training time per epoch is around 90 seconds. The validation error computation that’s run every epoch on the test dataset takes around 65 seconds. Thus total time per epoch is ~155 seconds (Full training takes roughly 1.3 hrs (30 epochs)).