A new approach for the prediction of partition functions using machine learning techniques

Area of Science:

• Thermodynamics
• Computational Chemistry
• Materials Science

Background:

• Predicting thermodynamic properties of atomic and molecular fluids is crucial for materials science and chemistry.
• Traditional simulation methods can be computationally expensive and time-consuming.

Purpose of the Study:

• To develop a machine learning (ML) approach for rapid and accurate prediction of fluid thermodynamic properties.
• To eliminate the need for extensive simulations in determining properties like Gibbs free energy, Helmholtz free energy, and entropy.

Main Methods:

• Training neural networks using results from flat-histogram simulations as a reference.
• Utilizing trained neural network weights to predict partition functions and thermodynamic properties.
• Validating ML predictions against experimental data and existing simulation results for argon, CO2, and water.

Main Results:

• ML predictions show excellent agreement with experimental and simulation data for thermodynamic properties.
• Highly accurate predictions for Gibbs free energy, Helmholtz free energy, and entropy across wide temperature and pressure ranges.
• The ML approach provides instant access to thermodynamic properties, including phase coexistence conditions.

Conclusions:

• The developed ML approach offers a computationally efficient alternative to traditional simulations for predicting fluid properties.
• This method accelerates the screening of new materials and the parameterization of force fields.
• Instantaneous access to thermodynamic data (G, A, S) facilitates faster scientific discovery and engineering applications.