A unified machine-learning framework for ab initio multiscale modeling of liquids

Area of Science:

• Physical Sciences
• Computational Chemistry
• Materials Science

Background:

• Predicting liquid matter behavior from microscopic interactions is a major scientific challenge.
• Accurate descriptions of intermolecular forces and multiscale frameworks are needed.

Purpose of the Study:

• To develop a computationally efficient and accurate framework for multiscale fluid modeling.
• To bridge the gap between microscopic interactions and macroscopic fluid properties.

Main Methods:

• Combined machine-learned interatomic potentials (MLIPs) with neural classical density functional theory (cDFT).
• Used MLIPs trained on quantum-mechanical data to generate density profiles for training neural cDFT.
• Applied the ab initio neural cDFT framework to water and carbon dioxide.

Main Results:

• The ab initio neural cDFT framework is more efficient than molecular simulations.
• Accurately reproduced bulk equations of state and phase diagrams for water and CO2.
• Predicted effects of confinement on water's liquid-vapor coexistence and supercritical CO2 behavior.

Conclusions:

• Established a general first-principles route to multiscale fluid modeling by unifying MLIPs and neural cDFT.
• Demonstrated a significant step towards generalizing cDFT for chemically complex systems.
• The framework provides a transparent route to fluid thermodynamics across scales.