Sarkas: A Fast Pure-Python Molecular Dynamics Code

Sarkas is a pure-python molecular dynamics (MD) code we are currently developing for wide (open source) use. To our knowledge, Sarkas is the first production-scale pure-Python molecular dynamics code designed for simulating plasmas. Current areas of application are strongly coupled plasmas which include dusty plasmas, ultracold neutral plasmas,  and warm dense matter.

Sarkas offers the ease of use of Python while being highly performant with execution speeds comparable to or exceeding that of compiled languages (e.g., C). Sarkas’s high-performance originates from extensive use of Numpy arrays and Numba’s just-in-time compilation.

Our work showed that MD simulations can be used in tandem with experiments to study the complex phenomena in high energy density plasmas. In this work we used Sarkas to study temperature relaxation in binary ultracold neutral plasmas. 

 

Here, we show how MD can be used to study kinetic instabilities in plasmas. In the movie below, the vertical axis is particle velocity and the horizontal axis is space (that is, it is phase space). The plasma is mostly composed of particles in equilibrium (red points) with a bump on the tail (blue particles at large, positive velocity). This situation mimics a plasma with an injected pulse of fast particles in it. What is interesting about this MD simulation is that it resolves particles down to their atomic scale, while successfully generating longer wavelength instabilities, thereby spanning extreme length scales. This shows promise for using Sarkas to study a wider range of kinetic and hydrodynamic phenomena building from the atomistic level. Read more in our paper in Phys Rev Research

 

 

As an example, MD simulations of strongly coupled plasmas enable access to a plasma’s static and dynamic properties. Static structural information is obtained through the radial distribution function g(r) using the particle positions. Dynamical properties such as transport processes  are also available by computing the transport coefficients from the particle trajectories. As a example, in the figure below the results of dynamic structure factor S(k,ω) for a Yukawa plasma, and the resulting wave dispersion properties.

 

Interparticle interactions currently included are long-range Coulomb and intermediate- to short-range screened Coulomb (Yukawa interactions), quantum statistical potentials, generalized Lennard-Jones potential. The computational cell can be either 3D or 2D with periodic, reflecting or open boundary conditions. For short range potentials, O(N) scaling is achieved with a neighbor list (linked cell); conversely, for intermediate- to long-range forces the Particle-Particle Particle-Mesh (PPPM) algorithm is employed.

To illustrate performance of Sarkas we compared the execution times with a equivalent version of the code written entirely in C. The figure below compares C and Python timings of the two most computationally intensive parts of the code: the Linked-Cell-List algorithm used in the Particle-Particle part and charge assignment to a grid for the Particle-Mesh part. The times (Intel Xeon E5-2670v2, 2.6GHz) are for one step of the force calculation for 10⁴ – 10⁶ particles. The performance of Sarkas is similar to that of the C version, even with C’s O3 compiler optimization. This is a remarkable result considering that Sarkas is a pure Python (with Numba) code.

 

You can find more details on the speed and capabilities used in Sarkas in our paper in Computer Physics Communications. Details of this the theoretical and numerical aspects can be found in this paper.

Some other projects can be found on Sarkas github repo