Skip to content

README.md

Latest commit

 

History

History
75 lines (56 loc) · 2.81 KB

README.md

File metadata and controls

75 lines (56 loc) · 2.81 KB

Forge

A HPC workflow management and materials analysis toolkit built with Python, ASE, and MACE.

Features

  • Adversarial Attack workflows for materials potential uncertainty
  • HPC job templates for VASP, MACE training, NEB, etc.
  • Composition analysis (short-range order, clustering, etc.)

Installation

Step 1: Install PyTorch and mace‑torch Manually

Forge does not include PyTorch or mace‑torch in its dependency list. Please install them manually based on your system's requirements. Additionally, you will need a CUDA-enabled GPU to run FORGE (and preferably slurm for spawning multiple jobs at once).

### Setting up forge conda/mamba environment
conda create -n forge python=3.11

# for cuda 12.6
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
pip3 install cuequivariance cuequivariance-torch cuequivariance-ops-torch-cu12

# for cuda 11.8
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip3 install cuequivariance cuequivariance-torch cuequivariance-ops-torch-cu11

# Optional for WANDB logging during training
pip3 install wandb

## Installing MACE 
git clone https://github.com/ACEsuit/mace.git
pip install ./mace[wandb] # ./mace without wandb

Step 2: Install Forge and dependencies

All of the needed dependencies are specified in the pyproject.toml file and can be installed with pip install -e . from the repository root.

## Basic installation of Forge
git clone https://github.com/mstapelberg/forge.git
cd forge

pip install -e .

Quick Usage

Refer to the scripts in forge/scratch/scripts or notebooks in forge/scratch/notebooks for examples.

Repository Layout

  • analysis/ - analysis scripts, e.g. Warren-Cowley
  • core/ - database manager, calculators
  • workflows/ - HPC job generation scripts, templates
  • tests/ - test suite
  • config/ - example database config
  • scratch/ - current scratch work repository

HPC Profiles

You can create profiles for clusters that you use. An example is put together for Perlmutter-CPU and Perlmutter-GPU. It is advised to have different types of profiles for jobs on different clusters. In the future we will implement folders for each cluster that you can save your job templates to.

Adversarial Attack Workflow

  1. Calculate variance: ...
  2. Run optimization: ...
  3. Generate Structures: ...
  4. Pick Structures for VASP: ...

Configuration

The central tenant of FORGE is using AWS relational databases. Configuration of your database is done in config/database.yaml. This allows you to just call DatabaseManager() without passing in a config file.

Contributing

We use ruff for linting and black for formatting. Contact myless@mit.edu if you'd like to contribute or drop an issue/discussion.

License

MIT License