Deep neural network learning of complex binary sorption equilibria from molecular simulation data

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Sorption equilibria are crucial for processes like desorptive drying.
• Molecular simulations accurately model these equilibria but are computationally intensive.
• Developing efficient predictive models is essential for process optimization.

Purpose of the Study:

• To develop a deep neural network (NN) as a surrogate for molecular simulations of complex sorption equilibria.
• To predict equilibrium loadings for alkanediol/water and alkanediol/ethanol systems in zeolites.
• To enable faster optimization of desorptive drying conditions.

Main Methods:

• Employed canonical (N1N2VT) Gibbs ensemble Monte Carlo simulations.
• Trained a multi-task deep neural network (NN) on simulation data.
• Utilized probabilistic modeling for sorption equilibria prediction.

Main Results:

• The NN accurately reproduced simulation results for various alkanediol/solvent/zeolite systems.
• The NN generated continuous isotherm functions, facilitating optimization.
• Learned representations in hidden layers enabled transfer learning for new systems.

Conclusions:

• Deep neural networks offer an efficient and intelligent surrogate for molecular simulations in predicting sorption equilibria.
• The developed NN model accelerates the optimization of desorptive drying processes.
• Transfer learning capabilities of the NN reduce data requirements for new system predictions.