Accelerating phase diagram construction through activity coefficient prediction.

Area of Science:

• Computational chemistry and thermodynamics.
• Machine learning applications in physical sciences.

Background:

• Predicting phase diagrams using molecular simulations is computationally intensive.
• Accurate thermodynamic data is crucial for understanding mixture behavior.

Purpose of the Study:

• To develop an efficient machine learning methodology for predicting phase behavior.
• To reduce the computational cost associated with phase diagram determination.
• To establish a predictive link between Kirkwood-Buff Integrals and activity coefficients.

Main Methods:

• Training a Gaussian process (GP) model on Kirkwood-Buff Integrals (KBIs).
• Utilizing KBIs to predict activity coefficients, which quantify deviations from ideality.
• Applying the trained GP model to new Lennard-Jones mixtures without prior phase data.

Main Results:

• Successfully predicted activity coefficients for new systems, bypassing direct coexistence simulations.
• Demonstrated significant reduction in computational expense for phase behavior prediction.
• Established a scalable framework for analyzing complex mixtures.

Conclusions:

• The developed machine learning approach offers a computationally efficient alternative for phase diagram prediction.
• This method is broadly applicable in computational thermodynamics for studying mixtures with tunable interactions.
• Leveraging KBIs with GP models provides a powerful tool for accelerating thermodynamic calculations.