(73b) Molecular Simulations and Machine Learning for Multicomponent Adsorption

Authors: 
Josephson, T. R., University of Minnesota
Sun, Y., University of Minnesota
Siepmann, J. I., University of Minnesota
Industrial catalytic and separation processes use complex, multicomponent feedstocks, and often operate at high temperatures and pressures. Consequently, process design requires knowledge of multicomponent adsorption across “process space” – the multidimensional space of temperature, pressure, and composition. Measuring fluid properties and adsorption under in situ conditions is challenging – especially when attempting to cover wide composition ranges of multicomponent mixtures.

We performed high-throughput Monte Carlo simulations in the Gibbs ensemble to predict real fluid properties of liquid, vapor, and supercritical fluids in equilibrium with zeolite adsorbents. We examined unary adsorption and competitive adsorption of mixtures with up to eight components. These simulations provide numerous discrete data points that are used to fit an artificial neural network, which provides a single, self-consistent, continuous, and differentiable function describing the properties of all fluids and adsorbed phases across the full range of realistic process conditions. Properties and their derivatives from the neural network can enable accurate predictions for process optimization and control.