Variational autoencoders for Lennard-Jones-like configurations

Directory structure

.
├── mc-sampling
│   ├── dump
│   ├── good-runs
│   ├── src
│   └── test
├── slides
└── vae
    └── saved-models
        ├── decoder
        └── encoder

Main notebook and vae-related files

The most comprehensive Jupyter notebook is located in the vae subfolder, named vae-training.ipynb. Within this folder, you can also find a stand-alone Python code that corrects the unwanted rotations introduced by the configuration reconstructor. Additionally, the saved-models folder contains several zip files of saved models from previous runs (one of which is unzipped and can be loaded directly from the notebook).

C++ code for Monte-Carlo sampling

To build the executable, navigate to the mc-sampling subfolder and run the make command. If you need to modify the run parameters, you can do so in the sampling.cfg file (or another file of your choice following the same conventions). After modifying the parameters, pass the configuration file to the executable using the command: ./main sampling.cfg.

The following parameters must be specified (order is not important). Use the exact parameter names followed by an equal sign without surrounding spaces and provide the corresponding value:

1. potential: specify the potential keyword along with its associated parameters (refer to the sampling.cfg file for more details).
2. n_parts: set the number of particles in the box.
3. side: define the side length of the box in units of sigma (Lennard-Jones parameter).
4. temp: specify the temperature in reduced Lennard-Jones units.
5. max_disp: set the maximum displacement per trial move in units of sigma.
6. n_systems: determine the number of separate initializations for the same systems. If set to 1, the program will generate an initial configuration once and continue to evolve that system. It will print the positions and optionally the energies at every n_sample timesteps (see below). By default, the program generates multiple systems and only prints the final configuration after n_steps timesteps.
7. n_steps: specify the total number of timesteps for each of the n_systems sub-runs.
8. n_sample: control the frequency of energy printing (if print_en = true) and position printing (if n_systems = 1).
9. print_en: a boolean value to choose whether to print energies or not.
10. show_z: print the average of the “last” coordinate (the second coordinate in two dimensions or the third coordinate in three dimensions) to the standard output.

To switch between two or three dimensions, manually change the macro

#define DIM 2

in the metropolis.hpp file located in mc-sampling/src to the desired dimension (2 or 3) and recompile the program.

The program generates two txt files, which are stored in the mc-sampling/dump directory. The files are named using a string that includes all the chosen parameters in the specified order, such as:

gamma_0.1_30_10_0.75_0.2_10000_2500_2500_x.txt

You can find some examples from past runs in the mc-sampling/good-runs directory. These examples are also referenced in the vae-training notebook for training and analysis.

Other stuff

You can use the format_all script to apply cosmetic edits to all Python files in the current directory and its subdirectories. To run the script, you need to install black by executing python3 -m pip install black, or conda install black if you are using Anaconda.