Tuning force field parameters of ionic liquids using machine learning techniques

The ability to estimate the force field parameters of materials is critical for calculating the properties of materials using molecular dynamics (MD) simulations. The density functional theory and empirical methods are used to calculate the force field parameters. These force filed parameters must be tuned in order to accurately capture the properties of materials using MD simulations. For performing MD simulations in ionic liquid systems, we tuned the force field parameters using Machine Learning techniques. We have specifically provided an example for tuning partial charges derived from the general Amber force field (GAFF) parameters for ionic liquids (ILs). We have performed equilibrium MD simulations with various partial charges at varied temperatures and pressures, and we obtained density of ILs at each temperature and pressure. We then use this MD simulation data to train several ML models to predict partial charges based on the experimental features of ILs. We employ several renowned techniques such as LASSO, Elastic Net, Partial Least Squares Regression, and Random Forest along with two stacked models using Mean Stacking and Linear Stacking. The predictions from our ML-based models follow the experimental data closely with the stacked models displaying the best overall predictive performance. The prediction model presented here is general and could be used to tune the force field parameters for a wide variety of application systems.

Automated summary

The ability to accurately estimate force field parameters for materials using MD simulations is crucial, and Machine Learning techniques have been employed to tune these parameters. The ML-based models show close prediction to experimental data, with stacked models performing the best overall predictive performance and providing a general approach for tuning force field parameters in various application systems.