Using Computationally-Determined Properties for Machine Learning Prediction of Self-Diffusion Coefficients in Pure...

Area of Science:

• Computational chemistry and materials science
• Physical chemistry

Background:

• Predicting liquid transport properties is crucial for understanding fluid behavior and designing new materials.
• Artificial Neural Networks (ANNs) have shown promise in predicting diffusion based on experimental data.

Purpose of the Study:

• To apply ANNs to predict the diffusion properties of pure liquids using data solely from molecular simulations.
• To identify key molecular and fluid properties that drive diffusion prediction accuracy.

Main Methods:

• Employed classical molecular dynamics (MD) simulations to obtain self-diffusion coefficients for 102 diverse pure liquids.
• Utilized fluid properties from MD simulations and molecular properties from quantum calculations as input features for ANNs.
• Performed feature sensitivity analysis to determine the most impactful input parameters for the ANN model.

Main Results:

• The MD-based ANN successfully predicted self-diffusion coefficients using only 2-3 key input features.
• Isothermal compressibility, heat of vaporization, and thermal expansion coefficient were identified as the most influential properties.
• A secondary ANN model using experimental data showed good correlation, despite limitations in data points.

Conclusions:

• Machine learning, particularly ANNs fed with simulation data, offers a powerful and efficient method for predicting liquid diffusion.
• This approach can significantly accelerate the understanding and design of materials and processes involving liquid transport.