Thermodynamically-guided machine learning for thermophysical property modeling and prediction

The modeling of thermophysical properties is central to chemical process design and optimization. These include thermodynamics and transport properties. Thermophysical properties have been modeled using empirical, semi-empirical, and theoretical approaches. Despite their success, current models rely on incomplete theories fitted to limited data. To address this issue, this thesis develops a hybrid approach combining molecular simulations, theory, and artificial neural networks (ANNs) for modeling thermophysical properties. This approach is demonstrated for systems interacting via the Mie potential. Thermodynamic properties can be determined from the derivatives of the Helmholtz Free Energy surface, which is modeled using a physics-embedded ANN Equation of State (FE-ANN EoS). This EoS is formulated to satisfy known thermodynamic constraints. The FE-ANN EoS is comparable to the monomer term of the theoretically derived SAFT-VR-Mie EoS. Moreover, the FE-ANN EoS provides a flexible framework for finding the appropriate functional form of the free energy, including contributions where an analytical solution is unavailable. The FE-ANN(s) EoS extends this framework to handle both fluid and solid phases. The resulting EoS predicts the entire phase diagram and provides insights into the continuity between different aggregation states. A similar physics-informed data-driven approach is used to model transport properties. In this case, ANN models are constrained by the dilute gas limit predicted by kinetic theory. These models accurately predict transport properties across a broad phase space region. The models developed in this thesis comprehensively characterize Mie particles. These models enable testing parametrization schemes for quasi-spherical molecules. The Mie potential exhibits a trade-off between fitting thermodynamic and transport properties. Incorporating transport properties or solid-phase data yields similar molecular parameters, suggesting that force field parametrization should account for this information. This thesis demonstrates how scientific machine learning enhances thermophysical property modeling, improving accuracy and addressing the limitations of conventional methods.