Bypassing traditional molecular dynamics with artificial neural networks

Abstract

An attempt has been made to speed up molecular dynamics simulations using machine learning. LAMMPS package was used to generate data for training the ML model which was programmed in PyTorch. The fidelity of the data generated by LAMMPS was first validated by simulating the evaporation of an Argon droplet in its own vapor. Results from simulations were compared with the D2 Law of droplet evaporation and a reasonably good agreement between theory and simulation was observed. Training and testing datasets consisted of per-timestep snapshots from 6 simulations of equilibration of up to 100 atoms in a periodic box. These were converted to images of dimension (10,10,6), such that 100 pixels of dimension (1,1,6) stored the coordinates and velocity components (x,y,z,vx,vy,vz) of up to 100 atoms. A Symplectic Recurrent Convolutional Hamiltonian Neural Network (SRCHNN) was proposed in which a conserved scalar analogous to the Hamiltonian of a system of interacting atoms was modeled using a Convolutional Neural Network. Using Hamilton's equations of motion, time derivatives of positions and velocities were obtained by taking the symplectic gradient of the Hamiltonian, calculated using backpropagation. Symplectic time integration with the Leapfrog algorithm was employed for predicting trajectories using the calculated time derivatives. The model was trained in a recurrent manner with sequences of particles’ positions and velocities. The performance of SRCHNN was tested against the length of the sequence used for training, ranging from 1 to 6. The mean square error (L2 loss) between the true and predicted output states did not decrease significantly with larger training sequence lengths. The percentage error between the predicted and true number of droplet particles was least for the smallest sequence length of 1; while the percentage errors between the droplet and ambient temperatures were roughly the same for all training sequence lengths. The SRCHNN was able to predict 15 future states in sequence within acceptable degrees of accuracy and a 3.1x speedup over LAMMPS was observed.