Figure 1.

This study considers the identification of structure–property relationships between a solute molecule and a target thermodynamic property. Rather than directly learning the relationship, we propose a three-step process: (1) molecular dynamics simulations of the solute in its condensed-phase environment; (2) structural analysis of the liquid structure; and (3) relating structural order parameters and thermodynamic properties. The learning procedure thus relies on features that incorporate relevant physics, including Boltzmann phase-space averaging, liquid environment, and collective effects. Our methodology is able to identify complex structure–property relationships, even with a simple linear model. Icon from flaticon.com.