Bayesian Optimization for Automated Molecular Dynamics Simulations

1. Clone the repository:

git clone https://github.com/PaulsonLab/BO4MD.git

2. Navigate to ~/BO4MD:

cd ./BO4MD

3. (Optional) Uncomment lines 19-27 in pyproject.toml if you have CUDA 12.4 on your system. This will enable installation of PyTorch with GPU support.

4. Install via poetry:

poetry install

Note: poetry needs to be installed first.

You can run Bayesian Optimization directly from the command line:

poetry run python -m bo4md --smoke_test true --acq logei --d 3 --n-init 10 --n-iter 20 --patience 5 --seed 42 --plot true --report true --outfolder ./out

• --smoke_test (bool, default: true)
  If true, runs the built-in synthetic smoke test function.
  If false, runs the MD simulator function (must be implemented by the user).

• --acq (str, default: "logei")
  Acquisition function. Choices: ucb, ei, logei, random.

• --d (int, default: 3)
  Input dimensionality (number of simplex components).

• --n-init (int, default: 10)
  Number of initial random samples drawn from the simplex.

• --n-iter (int, default: 20)
  Maximum number of Bayesian Optimization iterations.

• --patience (int, default: 5)
  Early stopping patience (stop if no improvement for this many iterations).

• --seed (int or None, default: None)
  Random seed for reproducibility.

• --plot (bool, default: true)
  If true, generates a plot of best-so-far objective value vs. iteration (bo_traj.png).

• --report (bool, default: true)
  If true, writes a text report of the optimization trajectory (out.txt).

• --outfolder (str or None, default: ./out)
  Output folder for initial data file, trajectory plot, and report.

Run a quick smoke test with default settings:

poetry run python -m bo4md

Run a smoke test with 30 iterations using UCB acquisition, 15 initial samples, and a fixed seed:

poetry run python -m bo4md --smoke_test true --acq ucb --n-init 15 --n-iter 30 --seed 123

Run with the MD simulator instead of the smoke test:

poetry run python -m bo4md --smoke_test false 

The md(.) function in simulator.py serves as a template which needs to be adapted to call real MD simulations. It expects an [n, d] torch tensor as input and returns an [n] torch tensor (simulation output), where n is number of data points and d is input dimension.

TBD