Machine learning for non-additive intermolecular potentials: quantum chemistry to first-principles predictions.

Area of Science:

• Computational chemistry
• Machine learning
• Physical chemistry

Background:

• Accurate potential energy surfaces (PES) are crucial for predicting thermophysical properties from molecular principles.
• Current methods for generating PES can be computationally expensive and require extensive data.

Purpose of the Study:

• To present a widely-applicable method for generating first-principles PES using quantum chemistry and machine learning.
• To demonstrate the method's accuracy and efficiency in interpolating three-body non-additive interactions.
• To apply the method to CO2-Ar mixtures and predict their thermophysical properties.

Main Methods:

• Utilized Gaussian Processes (GP), a machine learning technique, to interpolate three-body non-additive interaction data.
• Developed a method that requires no bespoke modifications for different molecular systems.
• Generated PES for CO2-Ar mixtures and calculated virial coefficients up to the 5th order.

Main Results:

• Achieved highly accurate interpolation from significantly fewer training points compared to traditional methods.
• Calculated CO2-Ar virial coefficients, leading to a virial equation of state (EoS) convergent up to critical density.
• Successfully predicted thermophysical properties like compressibility factor, speed of sound, and Joule-Thomson coefficient for CO2-Ar mixtures.

Conclusions:

• The GP-based method provides accurate, first-principles PES, enabling precise thermophysical property predictions for molecular mixtures.
• This approach reduces the need for extensive experimental data and computational resources.
• The method holds significant potential for applications requiring accurate continuum models and predictions for various molecular mixtures.