In this project, we study the frictional properties of a system with varying asperity locations on a 4x4 grid.
The procedure and consicderations can be found in the thesis paper master_final.pdf
All files needed to reproduce the thesis can be found here.
Both preprocessing, simulation launching, postprocessing and plotting is done in Python. The actual simulations are run using LAMMPS. Below, we have made a sketch of the structure of this repository. All actions are performed from the python directory: System setup is done by running the file python/pre/setup_system.py, launching simulations is done by running one of the files python/run/run_*.py, postprocessing dump files using Ovito is done by python/post/ovito_utils.py and plotting is done by the files python/plot/plot*.py.
├── initial_system
├── lammps
├── python
│ ├── pre
│ ├── run
│ ├── post
│ ├── plot
│ └── ml
│ └── fig
├── txt
│ ├── area_push
│ ├── area_relax
│ ├── coordination
│ └── max_static
├── fig
| ├── png
│ └── pgf
└── README.md
We preprocess the system using molecular-builder, which again is built around the ASE package. The asperity was prepared by carving out a rectangular octahedron from a 3C-SiC block with distance 11.19 nm between the (110) planes and the center of the asperity, and then a convex rectangular dodecahedron with distance 11.70 nm between the (111) planes and the center of the asperity. That makes the asperity similar to the equilibrium shapes of 3C-SiC found by Sveinsson et al. [1]. Then the asperity was cut and attached to the upper plate. The lower plate is just a (30nm x 30nm x 5nm) 3C-SiC block oriented either with the (100) or (110) plane pointing in z-direction. This process is repeated in a NxN (in our case a 4x4) grid, with the option for erratic disparity placements.
To launch the LAMMPS molecular dynamics simulations, we use the lammps-simulator package. We conscider two types of LAMMPS simulations: Relaxation simulations where the asperity is at rest and push simulations where the asperity is moved at a constant velocity. The simulations are run from python/run/run_relax.py, python/run/run_push.py 'python/run/run_relaxpush.py' and 'python/run/restart_push.py' respectively, which again call lammps/in.relax, lammps/in.push, 'lammps/in.relaxpush' and 'lammps/in.restartpush'. The simulations are computationally intensive.
We perform molecular dynamics simulations of the system in LAMMPS [REF], with the SiC force-field and parameters found by Vashishta et al. [REF] The uppermost and lowermost 1 nm of the system is held rigid (but the lower plate is allowed to move freely in the z-direction). A Langevin thermostat [REF] with damping time (??) is applied on the regions defined by z E [1, 2]nm U [18, 19]nm, which effectively sets the temperature of the system. The equation of motion is solved by a Verlet integration scheme [REF] with a time step of 2 fs and periodic boundary conditions. All simulations were run on NVIDIA A100 graphics cards using the KOKKOS package in LAMMPS.
The contact area analysis and coordination analysis are done using OVITO [REF]. The script for this is found in python/post/ovito_utils.py. The max static friction was found as the first prominent peak of the load curves when pushing the asperity. The contact area for the various simulations as a function of time is saved as text files in txt/area_relax or txt/area_push. A sigmoid curve is fitted to the load curves around the time of push. The rise of this slope denotes the stifness of the system, and the highets sigmoid value is an alternative to max static which to some degree counteracts noise in the data.
We use matplotlib [REF] to plot the graphs, and convert them to PGF-format for better rendering in LaTeX. To plot either the text files or LAMMPS log files, use the scripts in python/plot. Generated plots can be found in the fig directory.
Analysis of the resulting simulations are performed using pytorch. The object is to vary the asperity placement caracterized by a NxN boolean vector on the form [[1,0,1,0],[0,1,0,1],[1,0,1,0],[0,1,0,1]] where a boolean true denotes the location of an asperity. The output and the prediction is either the slope, the max static or the max simoid value. Both DNN and CNN networks are applied in large gridsearches over thousands of parameters using cross validations and other optimization strategies.
[1] Henrik Andersen Sveinsson, Anders Hafreager, Rajiv K. Kalia, Aiichiro Nakano, Priya Vashishta, and Anders Malthe-Sørenssen Crystal Growth & Design 2020 20 (4), 2147-2152 DOI: 10.1021/acs.cgd.9b00612