Speaker
Description
Polymorphism, the ability of molecules to form solid phases with slightly different arrangements, profoundly influences the material properties of molecular crystals in both pharmaceutics and organic electronics. Even minor variations in molecular packing can lead to vastly different charge transport properties in organic semiconductors. Indicative of this phenomenon are also subtle differences in the relative energies of polymorphic phases, often on the order of just a few meV. Accurately capturing these fine energy distinctions is challenging, as the uncertainty in machine-learned potentials (MLPs) can exceed the energy differences between polymorphs. We focus on systems such as quinacridone, a pentacene derivative, where the hetero-atom substitution facilitates categorically different hydrogen bonding motifs. Despite their unique structures, the energy variations between two of these polymorphs are as small as 0.1 meV/atom.
To model such systems, we are developing a workflow to train MLPs that can accurately capture these subtle energy differences to describe diverse polymorphic phases. Specifically, we are testing three training strategies for general MACE models: (1) naïve training on mixed datasets of three primary polymorphs, (2) training a multi-head model with cross-learning between polymorph-specific heads, and (3) finetuning foundation models to specialize in these systems.
The most effective model from this investigation is then applied to predict the thermal conductivities of three individual polymorphs using the Wigner transport framework. By comparing these predictions on equal footing, we aim to uncover how polymorphism and the accompanying structural variations impact heat transport in hydrogen-bonded molecular crystals. Additionally, this work provides valuable strategies for designing highly accurate MLPs for advanced modeling applications in organic electronics and pharmaceuticals.