Research Question: Do we achieve better performance with hyper-specified bond and angle parameters (and later clustering?)

Lily's Slides

In scope:

• 1 fit with Sage 2.2.1 valence
  • https://github.com/openforcefield/openff-forcefields
  • forcefield = ForceField("openff_unconstrained-2.1.0.offxml")
• 1 fit with parameters hyper-specified, w/o linearised harmonics
• 1 fit with parameters hyper-specified, with linearised harmonics
• Benchmarking QM

Nice to have:

• Incorporate neighboring atoms
• Fit with torsions over-specified
• Clustering final parameters
• Other benchmarks

Not in scope:

• Generating new data
• vdW fits

Getting Started

1. In home dir on UCI HPC3, clone this repo somewhere in /dfs9/dmobley-lab/user_id/ with:

   git clone https://github.com/openforcefield/back-to-school-jen.git

2. Install with:

	srun -c 2 -p free --pty /bin/bash -i
	cd back-to-school-jen
        CONDA_OVERRIDE_CUDA=11.8 mamba create -f environment.yaml --channel-priority flexible

Test that cuda is enabled with:

        mamba activate bts
        python -c "import torch; print(torch.cuda.is_available())"

3. From the 1_data directory get data from zenodo and reformat
4. From the 2_filtered_results directory filter out high energy conformations and any SMILES strings that cannot be parsed
5. Split filtered data into the training and test set from 3_split_train_test
6. (optional) Make a offxml file to fit in 4_make_offxmls
7. Create and save SMEE force field and topology inputs from openff interchanges in 5_setup_train_ff_topologies for each force field created in 4_make_offxmls.
8. In 6_run_fit run the fit using previously prepared files