The use of an interpretable machine-learning approach to disclose the molecular assemblage on melting transition.

contextThe access to the molecular assemblage on the melting transition is fundamental to phase-change material's design, particularly when applied to renewable and intermittent energy sources, and the investigation of physical/chemical/biological processes. The development of new methods to obtain this information is, consequently, highly desired. This framework is fertile ground for the establishment of machine learning-based models with both predictive and interpretable profiles. However, such models are difficult to establish. This work describes the implementation of a Random Forest interpretable approach set on the specific tree paths followed by a given chemical system (molecule) for the relationship between its descriptors/melting point value. The descriptors involve all combinations of atom pairs of a generical molecular chemical system. The combined use of Random Forest molecule's specific tree paths and descriptor concept enables the built model the capacity to highlight the most important combinations of pairs of atoms/interactions inherent to molecular assembly on melting stage. As proof of concept, this procedure was applied to investigate the organization of 2-(2,4-dichlorophenoxy)acetic acid (2,4-D) molecules at their melting point, with the results validated with thermo-regulated FTIR and computational chemistry approaches.

methodThis approach combines the Random Forest algorithm and atom-pair-based descriptor's pattern, set for a generical molecule, in order to find a straightforward structure-property relationship. The unique tree paths followed by a given molecule highlight new specific measures addressing a causality relationship involving its descriptors and the melting point profile. This machine learning approach was validated with thermo-regulated FTIR, interaction energies, and molecular dynamics techniques.