llms-for-lammps

Uses llms to design experiments to submit to lammps. Runs repeatedly to get toward some structure.

More details

See the messages variable in main to see what context the LLM is provided with to begin with, alternatively see the playground API here. The basic idea is that we've got some structure which was made by a melt-quench simulation, and we'll ask the LLM to try and infer the experiment (I've forgotten how it was made so). Obviously it's not given much context but we feed back the results to it and so maybe it can manage to infer the best parameters after a few datapoints.

See:

• main.py for hooking everything together. Mostly does the logic behind passing the response from the LLM and providing the next input.
• utils.py for; (i) formatting the details of the experiment (linking together all the necessary components of LAMMPS) (ii) functionality to submit a job to a SGE compute cluster (Sun-grid-engine) and await the result and (ii) analysis

Todo's

• More sophisticated organisation of all the inputs and outputs, its a bit of a mess at the moment
• Support OS LLMs
• Support other compute clusters, not just SGE
• Improve lammps/submit.sh to work on other clusters

Improvements:

• You can expand the analysis to feed back better information to the LLM

• You can change the way we get the LLM to infer the parameters of the simulation to lower the number of tokens required, lowering the cost of running the simulations continuously and avoiding context limits.

• You can give additional parameters to the LLM to infer to make things more impressive; change density, size, number of atoms etc in original structure.

Setup

pip install openai
pip install numpy
pip install quippy-ase
pip install python-dotenv

Also make sure you've got an installation of lammps in the immediate file directory (impossible).

Also change lammps/submit.sh as it may not be formatted to request resources properly for your cluster