Data-Driven Bayesian Optimization of P-Type Polymer Dopant Complexes for Next-Generation Thermoelectrics

Semiconducting polymers have shown remarkable performance in applications such as organic light-emitting diodes (OLEDs), organic solar cells (OSCs), and thermoelectric materials, establishing their status as formidable candidates for “greener” next-generation materials. These materials are solution-processable, hence their manufacture is low-cost and highly scalable. Doping suitable polymers with small molecules has a demonstrably positive impact on the electronic properties of the final polymeric thin film by increasing the charge carrier concentration and electrical conductivity. We have previously published several studies exploring different <i>p</i>- and <i>n-</i>type polymers, dopants, and their respective properties, showing, in some select cases, record-breaking thermoelectric performance. Unfortunately, the search for optimal polymer-dopant pairs thus far has been largely Edisonian and based on ill-quantified means like expert knowledge and chemical intuition. This work seeks to leverage the power of <i>ab initio</i> electronic structure calculations in conjunction with a machine-learning-based optimization for high-performing <i>p-</i>type doped semiconducting polymers.<br/>The goal of this work is to identify a novel high-performing <i>p-</i>type polymer-dopant combination from a combinatorically expansive pool of candidates, which will consequently demonstrate the utility of a data-based approach toward polymer engineering. Such an identification will be conducted using the Physical Analytics pipeLine (PAL), a Bayesian Optimization (BO) code written in Python and developed by the Clancy group. PAL has been shown previously to incorporate physical domain knowledge (<i>e.g.</i> DFT results) into an efficient optimization over a large compositional space. The input data will consist of electronic properties obtained from density functional theory (DFT) calculations, such as sigma profiles, HOMO-LUMO gaps, and ionization potentials/electron affinities. For reference, sigma profiles are unnormalized histograms of the screened surface charge of a molecule that provide important information about charge distribution in a molecule. This important information provides information on the degree of localization of charges, which impacts electron transfer. A large amount of model training data already exists from our previous papers about doped <i>p</i>-type polymers. The BO algorithm will attempt to maximize electron transport properties such as the Seebeck coefficient, power factor, and conductivity, in separate single-objective experiments. These properties will be calculated for candidate complexes using BoltzTraP2, a Python code that quantifies electron transport behavior.