In silico discovery of a new class of anolyte redoxmers for non-aqueous redox flow batteries

Abstract

Organic non-aqueous redox flow batteries (O-NRFBs) are emerging devices for storing intermittent renewable energy in the electric grid. For this application, redox-active organic molecules (redoxmers) are required that have suitable redox potentials, excellent solubility in electrolytes, and adequate stability in all states of charge. Due to the large available design space of redoxmers, machine learning is useful to identify optimal molecules that combine these properties. In this contribution, we propose a probabilistic algorithm that simultaneously expands structural diversity in a molecular library of redoxmer derivatives and limits it to synthetically accessible structures. A Bayesian optimization-based active learning algorithm is then used to discover promising molecules with a minimal number of computationally expensive quantum chemistry calculations. To demonstrate the power of this approach, we investigated derivatives of a redox active molecule, 2,1,3-benzothiadiazole. A library of 35 500 molecules was explored, and a new class of tricyclic derivatives with unusually low reduction potentials was discovered. We analyze and report the correlation between low reduction potentials, cyclic moieties, and positional specificity of functional groups. In addition, we report the electrochemical stability of selected molecules that display low reduction potentials and suggested molecules for the experimental validation of their promising electrochemical properties.