Learning Universal Fundamental Relations for Predicting the Phase Behavior of Liquid Mixtures

Predicting properties of liquid mixtures
is central to
chemical
engineering, yet accurate predictions from unimolecular properties
have proven elusive. Activity models are a popular class of mixture
models due to their parsimony yet strong correlative ability. The
latest evolution in activity models is data-driven machine learning
models using neural networks (NN). However, NN models require a substantial
amount of training data to make their predictions, and many lack an
accessible free energy function, hindering extrapolation. This work
presents a joint data- and theory-driven model learning a set of universal
fundamental relations (UFR) governing small molecule liquid mixtures.
UFR models accurately correlate experimental infinite dilution activity
coefficient data (IDAC). Trained only on IDAC data, UFR models can
predict T–x–y and P–x–y vapor–liquid equilibrium curves of mixtures outside
the training set. Several UFR models were also able to simultaneously
predict binary liquid–liquid phase separations.