Predicting long-term trends in physical properties from short-term molecular dynamics simulations using long...

Area of Science:

• Materials Science
• Computational Chemistry
• Artificial Intelligence

Background:

• Molecular dynamics (MD) simulations are crucial for understanding material properties but are computationally expensive.
• Predicting long-term material behavior from limited simulation data remains a challenge.

Purpose of the Study:

• To develop a novel deep learning model for accurate prediction of physical properties using partial MD simulation data.
• To significantly reduce the computational cost associated with predicting material properties over extended timescales.

Main Methods:

• Utilized a graph convolutional network (GCN) to extract latent vectors from atomic coordinates in MD simulations.
• Employed a long short-term memory (LSTM) network to learn temporal trends from these latent vectors.
• Integrated fully connected layers and residual connections for one-step-ahead prediction of physical properties.

Main Results:

• Achieved accurate one-step-ahead prediction of potential energy variations during Ni solidification and melting.
• Successfully enabled long-term predictions (over 900 ps) from initial simulation snapshots.
• Captured key physical phenomena, like solidification completion, not evident in short-term data.
• Reduced computation time by a factor of 700 compared to full MD simulations.

Conclusions:

• The proposed LSTM-based model efficiently predicts physical properties from partial MD data.
• This approach offers substantial computational savings for materials property prediction.
• The model demonstrates potential for accelerating materials discovery and characterization.