Equivariant learning of a transferable three-dimensional classical density functional

What the math gives to ML

The paper offers a transferable, variationally consistent way to learn a three-dimensional thermodynamic functional from equilibrium density fields rather than supervised free-energy or chemical-potential labels. Its key transferable mechanism is to represent an equivariant scalar functional of a spatial field, then obtain predictions and responses by functional minimization and differentiation, preserving the variational structure that generates phase behavior and forces. In neural networks, this suggests replacing direct field-to-field predictors with an energy functional whose Euler-Lagrange minimizer is the output, while using symmetry and Hessian diagnostics to control stability. The strongest falsifiable signature is that one learned functional should transfer across temperatures, system sizes, boundary conditions, and ensembles, with response quantities determined by the same functional rather than separately trained heads.