The Green-Kubo method is an approach to compute the thermal conductivity of materials, in particular those with highly anharmonic potential-energy surfaces, through equilibrium molecular dynamics simulations. In two recent papers, we introduced an approach to use "modern" machine-learning potentials which use message passing and automatic differentiation for this method.
This repository provides a minimal implementation of the GK method with jax, based on vibes (for the actual GK functionality, i.e., computing thermal conductivities), stepson (datasets), glp (heat flux and MD), and mlff (for the so3krates potential). The high-cost parts of the method, long-running NVE molecular dynamics, are executed entirely on the GPU. For convenience, we also have a wrapper to the ase Langevin thermostat for thermalising systems.
This is early-stage code. It works, but it is not yet optimised for ease of use. However, a new version is in development, which features a redesigned input format, logging, reduced output to save disk space, and on-the-fly heat flux computation every n-th step. Together, they make production calculations much more convenient. If you are considering gkx for production use, please get in touch to test the new version!
Nevertheless, look at it go:
This runs 0.1ns of MD with 512 atoms with the LJ potential. With a so3krates potential, speed is reduced by about an order of magnitude, so let's not get too excited!
The most difficult part will be getting the dependencies to run: You need a CUDA-ready install of jax, and then glp and mlff. Once that's done, pip install . in a clone of the repository should suffice, which will install the remaining dependencies.
gkx is currently extremely minimalistic. It has exactly two CLI commands:
gkx run md input.yamlloads instructions frominput.yamland then runs MDgkx out gk trajectory/performs the Green-Kubo processing and emits agreenkubo.ncdataset with the result
For additional information, until docs are written, please refer to the CLI itself or, even better, the code, for further details.
Input files look like this:
nve:
# timestep in fs
dt: 4.0
# number of timesteps
maxsteps: 250000
# size of written output chunks (must divide maxsteps)
chunk_size: 25000
# number of steps performed on the GPU in one call
# may need to be adapted to match available RAM
batch_size: 25
files:
# starting configuration (FHI-aims input format)
geometry: geometry.in
# "pristine" supercell before thermalisation (optional)
supercell: geometry.in.supercell
# primitive cell (optional)
primitive: geometry.in.primitive
# matching to the glp.instantiate conventions:
potential:
lennard_jones:
sigma: 3.405
epsilon: 0.01042
cutoff: 9.0
onset: 8.0
calculator:
atom_pair:
heat_flux: True
To run thermalisation, the nve block is replaced with
nvt:
temperature: 10
dt: 4.0
friction: 0.02This simply uses the ase implementation of the Langevin thermostat.
We defer to ase and vibes for units. Internally, everything is based on ase, so the timestep is internally in ase.units.fs rather than SI fs. The datasets that are written respect the FHI-vibes conventions, which use ase for everything except the heat flux, which is written in SI units and uses ps (rather than fs) as the base time unit. Stress and heat flux are divided by the system volume.