A Bayesian approach providing design choices and chemical insight for solution-processed thermoelectric polymers

Abstract

The advent of machine learning approaches to materials design is poised to \textit{lead} experimental validation by selecting a m uch smaller, more promising, and sometimes unexpected, set of new materials candidates. This work describes one such investigation to characterize novel, hypothetical, and promising polymer candidates based on their calculated molecular-scale electronic and chemical parameters. Subsequently, we optimize them for a chosen chemical property using a chemically informed Bayesian surrogate model. As a test case, over 7300 combinations of novel, hypothetical semiconducting diketopyrrolopyrrole-based (DPP) polymers and commonly used solvents, generated by density functional theory, were screened based on their free energies of solvation, G$_{solv}$, as a guide to selecting optimal processing conditions. From this synthetic data set, we trained a physics-informed Gaussian process model that linked molecular-scale electronic structure properties to G$_{solv}$, and then used Bayesian optimization (BO) to identify key descriptors for predicting optimal enthalpic, single-chain polymer solvation in the infinitely dilute limit. As a result, we predicted an ``optimal'' solvent dielectric constant value around 10 for the DPP-based polymer class. To validate this result, we showed that the predicted polymer design associated with the minimum $\Delta$G$_{solv}$ value corresponded to the polymer with the highest experimentally measured conductivity in the literature. The importance of these observations is that our BO approach provided the chemical insight necessary to quickly screen solvents for potential DPP-based polymers based on the polymer repeat unit's quadrupole moments and to identify preferred compatible solvents corresponding to a minimized $\Delta$G$_{solv}$. This study also highlights the effectiveness of our BO algorithm: The optimal polymer design was found using just 6\% of the parameter space and in a very short execution time (on the order of minutes), a feat which cannot be duplicated experimentally. Finally, to show the extensibility of this machine learning approach, we repeated this exercise with a second class of semiconducting polymers in which the DPP base group is replaced by an indacenodithiophene (IDT) group. We again successfully validated our machine learning predictions of the most promising polymer designs against experimental results.