RAVE is a machine learning-based framework that learns the reaction coordinates, which can be used in biased MD simulation to enhance the sampling. Please read and cite these manuscripts when using RAVE: https://aip.scitation.org/doi/abs/10.1063/1.5025487 https://www.nature.com/articles/s41467-019-11405-4 https://arxiv.org/abs/2002.06099
Overviews of each of the subdirectories in this repository are provided below. For detailed descriptions of the files within the subdirectories, refer to the README's within each subdirectory.
This file contains the main part of RAVE that uses a deep neural network to learn the reaction coordinates as linear functions of order parameters. The linear combination coefficients are written to a txt file.
This file contains the data preprocessing part of the code. It reads the output of MD simulation and converts them into the format that can be read by P_rave.py. It will be called by P_rave.py automatically.
This file contains the output analyzing part of the code. It reads the output of neural network and outputs the combination coefficients. It will be called by P_rave.py automatically. It is also useful to run this file separately in the situations that the main code is still running or stop before finishing due to running out of time on HPC.
Usage:
python P_rave.py
All the parameters need to be claimed correctly inside the file.
Some important parameters include:
-system_name : the name of system. Input trajectories files are named as: system_name_<traj_index>, where <traj_index> is the index of a trajectory starts from 0.
-n_trajs: the number of trajectories
-bias: whether the trajectories are from biased MD.
-time_delay: predictive time delay.
-T: temperature in unit of Kelvin.
-op_dim: dimensionality of order parameters.
-rc_dim: dimensionality of reaction coordinates.
-training_size: the number of data points for training. Training size should be smaller than the total number of available data points.
-batch_size: number of data points in each batch. The total number of training data point n should be a multiple of batch_size.
-epochs: number of epochs to train the model. A rule of thumb is training_size*epochs/batch_size should be around 10^5.
By default, input trajectories are in the input directory and the output can be found in output directory. Users should create these two directories before running the code.