Self-diffusion of liquids characterizes the random translational motions of molecules under thermal agitation. Predicting the self-diffusion coefficients of pure liquids is indispensable in liquid science, material physics, chemical engineering, and other topics in theoretical and applied chemistry. However, accurate self-diffusion coefficients of liquids are scarcely reported despite their importance.<br/><br/>The authors of the present study compiled a new large dataset containing over 1000 logarithmic values of the self-diffusion coefficients of chemically diverse pure liquids. The dataset combines the data of a literature survey with the experimental data of pulsed-field gradient nuclear magnetic resonance techniques.<br/><br/>Subsequently, a number of computational molecular descriptors were generated based on the two-dimensional chemical structures of each liquid and then the authors applied promising nonlinear machine learning algorithms such as support vector regression and random forest combined with stepwise forward descriptor selection. The resulting prediction models achieved excellent predictive performance. The externally validated determination coefficients and root mean square error between the experimental and predicted values were above 0.90 and under 0.30, respectively. Furthermore, the models satisfied the recently proposed statistical acceptance criteria for external validation and Y-randomization test of prediction models. Therefore, the models can be practical tools for high-throughput virtual assessments of liquid self-diffusion coefficients.

Quantitative prediction of self-diffusion coefficients of diverse pure liquids by machine learning techniques

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.