Description of transport properties of pure Mie Fluids using Machine Learning

Abstract

Transport properties of fluids are heavily dependant on the interactions between the fluid particles. These interactions are complex and require different models for each fluid to obtain an accurate description. An accurate model for the transport properties based on a family of variable interactions would allow for fast material screening and accelerate the material development cycle. In this work the relationship between the transport properties and a family of Mie particle interaction potentials is studied. In the first part of this thesis, a system of triangular trimers interacting with four different potentials is studied. These potentials have different values of the cohesive parameter resulting in observable differences in thermodynamic and transport properties. The critical temperatures decreased for a decreasing cohesive parameter, in agreement with previous findings for single sphere systems. The amount of anomalous diffusion was correlated with the cohesive parameter as well, with an increase in non-Gaussian dynamics for lower values of the cohesive parameter. In the next section we further explored the relationship between the transport properties and the Mie potentials. We developed three Machine Learning (ML) algorithms that could accurately predict the value of the self-diffusion coefficient and viscosity in a large range of state points with an $R*2>0.95$ and AARD$\leq 6\%$ for most methods. The ML models were then used to predict the transport properties of real fluids. The predictions were very accurate for low density states, shown by an $R^2>0.9$. Due to the simplicity of the proposed model system the observed AARD$\approx25\%$, showing that the high density region was not accurately represented by the ML algorithms. In the final section we use our previously obtained models to calculate the transport properties of a larger selection of fluids. The fluids were modelled based on their thermodynamic properties, and the obtained models used to calculate the transport properties. The ML models performed similarly to our previous study, giving a good description for the low density state points, but not high density as observed in the prior section. Depending on the strength of inter-particle interactions the ability of the models ability to predict the transport properties varied. We found that the largest contributor to errors in predictions was the presence of hydrogen bonding or anisotropy. Additionally, we produced models for the fluids based on their transport properties and found that this increased the accuracy for the high density region at the expense of the low density region showing decreases in both $R^2$ and AARD. The differences between the accuracies of the models were not large and any of the models could be used to a similar degree of accuracy in transport property prediction.

Date of Award	1 Aug 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Carlos Avendano (Supervisor) & Alessandro Patti (Supervisor)