DescriptorDOS 0.1.1

Score-matching the descriptor density of states for model agnostic sampling

DescriptorDOS is a scheme to return phase free energies as smooth functions of MLIP parameters, for uncertainty quantification in forward propagation and inverse fine-tuning in back-propagation.

Details are provided in our paper:
Score matching the descriptor density of states for model-agnostic free energy estimation
Preprint:arXiv 2025

DescriptorDOS is designed for the wide variety of MLIPs that are linear in some set of descriptor features.
As we discuss in the paper, whilst this clearly includes ACE/SNAP/qSNAP/POD/MILADY "as-is", multiple applications to non-linear architectures are envisioned and will be available in the near future.

Further applications / demonstrations of DescriptorDOS are coming- watch this space!

Quick overview

See below for installation (mpi4py and lammps), and examples/ for detailed usage.

DescriptorDOS requires a LAMMPS input file which reads in the atomic structure, loads the reference potential energy function and defines a global descriptor compute, e.g. SNAP

# The descriptor compute
compute D all sna/atom 4.7 0.99363 8 0.5 1
compute aveD all reduce ave c_D[*]

The compute name is then passed by a YAML file or as an argument: Via configuration.yaml:

lammps_input_script: "configuration-files/in.lammps"
global_descriptor_compute_name : "aveD"

At initialization

from mpi4py import MPI
from DescriptorDOS import Manager 
DDOS_Manager = Manager( comm=MPI.COMM_WORLD,
                         global_descriptor_compute_name="aveD", # overwrites YAML
                        yaml_file="configuration.yaml")
DDOS_Manager.run()

Installation

DescriptorDOS uses LAMMPS and mpi4py, in addition to numpy and scipy. If you can run

from mpi4py import MPI
from lammps import lammps
lmp = lammps(comm=MPI.COMM_WORLD)
lmp.close()

Then you can install DescriptorDOS:

pip install DescriptorDOS

Scripts in testing/unittests.py to check everything is set up correctly. Check out examples/ for a simple application. For HPC use we recommend installing lammps and mpi4py as described here

For local testing you can install LAMMPS via @jan-janssen conda-lammps (one CPU/worker)

conda config --add channels conda-forge # add conda-forge channel
conda create -n ddos-env python=3.10 
conda activate ddos-env 
conda install mpi4py lammps
pip install DescriptorDOS

We emphasize conda-lammps will not give optimal performance on HPC!

Algorithm Overview

The DescriptorDOS algorithm leverages MPI parallelism for efficient sampling and tensor-compression for low-rank storage.

Running Descriptor DOS

See examples/. Whilst sampling is fully parallel, to generate samples with Hessian displacer, we require the reference system's Hessian.

Parallel batching of these tasks is simple but current implementation simply runs in serial for 5-10 minutes before batch submission (see examples/run.py):

python run.py # will store Hessian data and only run a few samples

We can then run the DDOS sampling in parallel:

mpirun -np ${NPROCS} python run.py

Whilst we provide many options in run.py, they can all be stored in a yaml file, so the python script can be as succinct as as snippet (first seen above)

from mpi4py import MPI
from DescriptorDOS import Manager 
DDOS_Manager = Manager( comm=MPI.COMM_WORLD,
                        yaml_file="configuration.yaml")
DDOS_Manager.run()

Analysing data

We provide a sample notebooks in analysis/ of our "world-first" result, inverse fine-tuning of a phase transition temperature.

More examples will be added in the coming months.

(c) MIT 2024 TD Swinburne [e-mail]