Forecasting molecular dynamics energetics of polymers in solution from supervised machine learning | JoVE Visualize

Area of Science:

Computational Chemistry and Materials Science
Machine Learning Applications in Science

Background:

Machine learning, particularly neural networks, is increasingly used for analyzing chemical and physical systems.
Predicting kinetics and dynamics using machine learning remains less explored compared to structure and energetics.
Recurrent neural network (RNN) architectures are suitable for time-series data common in molecular simulations.

Purpose of the Study:

To evaluate the performance of three RNN architectures (ERNN, LSTM, GRU) in reproducing and forecasting the energetics of a macromolecular aggregate in solution.
To develop an improved in silico protocol for more reliable energetics forecasting in complex molecular systems.

Main Methods:

Training and testing ERNN, LSTM, and GRU models on a large time series dataset of potential and interaction energies from molecular dynamics simulations.
Developing a novel in silico protocol involving extraction of time patterns and ensemble modeling to enhance forecasting accuracy.
Analyzing the ability of models to reproduce time series and forecast energetics with appropriate statistical distributions.

Main Results:

The investigated RNN architectures excellently reproduced the time series data of macromolecular aggregate energetics.
Forecasting capabilities for short-term and long-term energetics were limited, showing deviations in statistical distributions.
The proposed in silico protocol significantly improved energetics forecasts, generating a band of time series consistent with simulation fluctuations.

Conclusions:

While standard RNNs effectively model current energetics, an ensemble-based in silico protocol enhances their forecasting utility for solvated macromolecular aggregates.
This data-driven protocol offers valuable insights into the system's future energetics, aiding decisions on simulation continuation.
The approach expands the application of supervised machine learning as a decision-making tool in scientific simulations and materials design.