Data‐driven approximation of thermodynamic phase equilibria

Abstract

We present a new data‐driven approach for both accurate and computationally efficient approximation of vapor liquid equilibria (VLE) models. Our method is able to provide guaranteed enclosure to limit the approximation errors over the entire domain of interest, all just by sampling only at select points. The approximation relies on a mixed‐integer linear programming (MILP) formulation that exploits vertex polyhedral properties of theoretically guaranteed lower and upper bounds to enclose nonlinear and nonconvex equations of state (EOS) and empirical models. Another advantage is that, unlike traditional full simulation‐based data‐driven approaches, we do not solve nonlinear system of equations (f(x) = 0) for sampling. Instead of looking for only feasible samples, we evaluate f(x) over x‐domain. This functional evaluation eliminates the need for computationally‐demanding full‐scale simulations and the associated convergence issues. We demonstrate excellent performance of the proposed MILP formulation in predicting the solubility of hydrofluorocarbon (HFC) refrigerants in ionic liquids (IL).

Separation Process Synthesis for High-GWP Refrigerant Mixtures: Extractive Distillation using Ionic Liquids

“…Since the accuracy of the process models is dictated by the rigorousness of the thermodynamic models, one can employ EoS-based solubility models (e.g., Peng−Robinson or RK) to obtain a more realistic process performance prediction at a higher computational time. To that end, a data-driven approach 74 offers exciting prospect to achieve optimal solutions fast, while maintaining the predictive capability of such rigorous models. Also, rather than using the equilibrium based model, one can use rate-based models considering transport properties with appropriate mass transfer-coefficient, viscosity, and other thermophysical properties (e.g., conductivity).…”

An Overview of Computer‐aided Molecular and Process Design

“…Data-driven surrogate [206][207][208] and ML models can provide a highly accurate prediction that is on par or superior to the heavily used thermodynamic equation of states. Also, the exact MILP formulation of the trained ReLU-ANN model allows for the development of a part of the molecular model as MILP.…”