PND: Physics-Informed Neural-Network Molecular Dynamics

We have developed PND, a differential equation solver software based on physics-informed neural network (PINN) for molecular dynamics simulators. Based on automatic differentiation technique provided by PyTorch, our software allows users to flexibly implement equation of motion for atoms, initial and boundary conditions, and conservation laws as loss function to train the network. PND comes with a parallel molecular dynamic engine in order to examine and optimize loss function design, and different conservation laws and boundary conditions, and hyperparameters, thereby accelerating PINN-based development for molecular applications.

Documentation

For documentation please visit this page

Code Capsule

Visit the code capsule here (https://doi.org/10.24433/CO.9413795.v1)

Build Command

Use CMake to build the target pnd_example. This is binary for predicting energies using auto gradient using hamiltonian. cmake --build . --config Release

Input Parameters

All system input parameters go into pmd.in

• The first line is vproc[3]
• The second line is InitUcell[3]
• Density
• Initial Temperature
• DeltaT or time step
• Number of warmup steps for MD - Results from these steps will not be considered for any trainig purpose. The purpose of this step to bring the system to active state
• Step Average - This is the number of steps to perform MD for before printing the state of the system to console
• Number of steps to perform pre-training
• Number of steps to perform main-training after pre-trainig
• Epochs for pre-training
• Epochs for main-trainig

The following is an example

1 1 1   # Cell size, i.e., dimensions of the parallel-piped system along (x,y,z)       
2 2 2   # Spatial decomposition, i.e., number of subsystems that will be assigned
          to MPI processes (l.b.h)
0.18    # Density or inter-atomic spacing
0.7     # Initial temperature
0.01    # The shortest time step for which trajectories must be computed
130     # Time at which the system reaches its relaxation state
1       # Interval to display output log 
25      # Time steps to predict 
50000   # Training epochs for pre-training
200000  # Training epochs 

Build on USC HPC

build_run_PND_Discovery.sh script can be used to build on USC HPC

Alternatively, the following commands can be used. You will first need to load the modules before building and run the executable with the following commands.

module purge
module load usc
module load cuda/10.1.243 python/3.7.6 cmake/3.16.2 cudnn/8.0.2-10.1

python3 -c 'import torch ; print(torch.utils.cmake_prefix_path)'

CC=gcc CXX=g++ 
cmake -DCMAKE_PREFIX_PATH='/spack/apps/linux-centos7-x86_64/gcc-8.3.0/python-3.7.6-dd2am3dyvlpovhd4rizwfzc45wnsajxf/lib/python3.7/site-packages/torch/share/cmake;/usr/lib64' .

cmake --build . --config Release

mpirun -np 1 -quiet ./pnd_example./pnd_example

Build note on Intel devcloud

cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=dpcpp -DCMAKE_PREFIX_PATH=$PWD/../libtorch/ 

and edit CMakeFiles/grad_lap.dir/flags.make

# compile CXX with /opt/intel/inteloneapi/compiler/latest/linux/bin/dpcpp
CXX_FLAGS =   -D_GLIBCXX_USE_CXX11_ABI=0   -D_GLIBCXX_USE_CXX11_ABI=0  -Wall -Wextra -Wno-unused-parameter -Wno-missing-field-initializers -Wno-write-strings -Wno-unknown-pragmas -Wno-missing-braces -openmp -std=gnu++14