Repository logo

Physical properties of alkanes and their mixtures



Change log



Alkanes and their mixtures are some of the physically simplest molecules and are widely used in industry, yet the connection between their structure and physical properties is still poorly understood. To make progress, we study the properties of pure alkanes with neural networks and molecular dynamics, while we develop a new theoretical framework to study the properties of mixtures of alkanes.

We first encode alkanes’ structure into five non-negative integers and use them as neural network inputs. Then, we utilize the neural networks to study the boiling point, vapor pressure, heat capacity, and melting point of light alkanes, as well as flash point and kinematic viscosity of linear alkanes. Neural networks model all these properties more accurately than the competing statistical and physico-chemical methods, while the cross-validation results indicate that they can confidently and accurately extrapolate the boiling point, heat capacity, and vapor pressure models to heavy alkanes. Still, due to a lack of experimental data for non-linear alkanes, neural network flash point and kinematic viscosity models cannot extrapolate to heavy alkanes, while the comparatively low accuracy of melting point models relative to other properties’ models suggests that additional physical effects need to be incorporated into them.

To obtain synthetic data as a supplement for the experimental kinematic viscosity dataset, we perform molecular dynamics simulations for density and non-equilibrium molecular dynamics (NEMD) simulations for dynamic viscosity. Density simulation results are corrected through a data-driven approach to increase their accuracy, and we develop a sampling algorithm that automatically selects the shear rates at which to perform the viscosity simulations.The sampling algorithm is tested on linear alkanes, and simulation results are in excellent agreement with the experiments, encouraging applications to more complex alkanes.

Then, we use neural networks with molecular structure as inputs to model the molecular dynamics density simulation values and extrapolate to 11-heptyltricosane, 8,11-dipentyloctadecane, and 8,14-dipentylhenicosane at 40°C and 100°C. These extrapolated density values are used as state points for the NEMD viscosity simulations, which are performed with the help of the shear rate sampling algorithm. While the accuracy of neural network models is high, and the usefulness and reliability of the sampling algorithm is further established, viscosity simulation results are not in a good agreement with the experiment due to systematic error in the force field.

Finally, to model properties of mixtures of alkanes, we develop a theory of mixtures whose molecules’ positions have a uniform spatial distribution. We apply this theory to molar volume, isentropic compressibility, surface tension, and dynamic viscosity of mixtures of alkanes, first by fitting to experimental data, and then by using the best fit parameters for viscosity to predict viscosity of further mixtures. Best fits and predictions show excellent agreement with the experiments, and our theory shows promise for further applications to mixtures of alkanes, while its conceptual basis has the potential to be applied to other types of mixtures as well.





Conduit, Gareth


Molecular Modelling, Molecular Dynamics, Neural Networks, Machine Learning, Statistical Physics, Alkanes, Physics, Chemical Physics, Numerical Optimisation, Algorithms, Simulations


Awarding Institution

University of Cambridge