2025 AIChE Annual Meeting

(649a) Universal Fundamental Relations for the Prediction of Thermophysical Properties

Authors

Bradley Olsen - Presenter, Massachusetts Institute of Technology
Oliver Xie, Massachusetts Institute of Technology
Phase behavior and thermodynamic property prediction are foundational tasks across Chemical Engineering, from designing more efficient separations to the discovery of battery materials to food engineering. The value of these predictions has motivated large efforts at both data collection and modeling, leading to the advanced prediction software that we use today. However, classical thermodynamic models must be regressed to data for the property of interest, making prediction in new systems challenging. The widespread adoption of machine learning offers the potential for improved accuracy, but models generally do not extrapolate from one data type to another and are also incapable of predicting unseen mixtures or properties. Inspired by universal differential equations (UDEs) from fluid mechanics that combine the underlying physics-based symmetries of systems with coefficient training on a diversity of data sets, we propose universal fundamental relations (UFRs) as a method for thermodynamic property prediction. We demonstrate how UFRs can be composed from classical, parsimonious mathematical forms for activity relationships, and thermodynamic data is then used to train a molecular embedding analogous to the Hansen solubility parameters that enables computation of activity from these relationships. Since this thermodynamic data can be conceptualized as a network in D-dimensional space, overfitting is easily avoided by requiring sufficient data to over specify this network for each individual node. Because the molecular embeddings are single-molecule properties, UFRs can easily extrapolate from a set of known mixtures to unknown mixtures as long as both molecules were represented within the overall training data. We demonstrate that the UFR approach allows prediction of infinite dilution activity coefficients (IDACs) with state-of-the-art accuracy. More impressively, the UFRs fit to only IDAC data can accurately predict vapor-liquid equilibrium (VLE) data and can semi-quantitatively predict liquid-liquid equilibrium (LLE) data without any additional measurements. Therefore, UFRs provide an accurate and extrapolatable method for thermodynamic property prediction that can be trained on sparse data to yield a single molecule-based parameterization, yielding an easily applied model for chemical engineering design.