Equivariant learning of a transferable three-dimensional classical density functional

Imagine you're trying to predict how liquids behave under different conditions, like temperature changes or when confined in small spaces. Currently, this often requires running separate simulations for each scenario, which can be time-consuming and inefficient. This is problematic because it limits our ability to quickly understand and manipulate liquid behavior in practical applications, such as in materials science or chemical engineering. This challenge is known as the need for separate atomistic simulations for each state, which can be cumbersome and resource-intensive.

To address this, the authors propose a method that learns a free-energy functional directly from three-dimensional equilibrium density fields. This approach allows for a single learned functional to be applicable across various temperatures and system sizes, effectively capturing the essential thermodynamic properties without needing specific training targets. By applying this method to complex geometries, they can predict behaviors like the forces involved in the formation of solvent-depleted bridges between colloids. This advancement means that builders can now leverage a more efficient way to connect microscopic liquid structures to their macroscopic behaviors, streamlining the process of understanding and designing liquid systems.

The approach introduces a novel way to learn thermodynamic functions from 3D density fields.

The results are validated across various conditions, though some aspects lack direct training targets.