A Radical Approach to Machine Learning and Electronic Structure Theory

Recent years have seen an explosion of interest in organic radicals due to their high efficiency in organic light-emitting diodes (OLEDs) and for use in qubits. However, modelling their electronic structure is challenging due to the large size of the molecules, their multireference electronic structure, and the spin contamination problem which can arise from multiple unpaired electrons in the excited states. In this presentation I will present ExROPPP, which we believe is the fastest known method for radical excited state computation, has accuracy comparable to high-level multiconfigurational calculations, and in spin-pure [1]. In addition, alternacy or pseudoparity rules for ground-state closed shell radicals had been derived in the 1950s by Pariser, Parr and Pople but there was no extension to radicals due to the algebraic complexity. In deriving ExROPPP we (Green and Hele) also derived alternacy or pseudoparity rules for the excited states of hydrocarbons radicals, showing their excited states also have 'plus' and 'minus' forms [1]. I will then show what we believe is the first example of machine learning directly from the excited states of radicals where we extend ExROPPP to molecules also containing nitrogen and chlorine by learning the optimal parameters from experimental absorption data and molecular structures [2]. These breakthroughs pave the way for the high throughput discovery of the next generation of radical-based optoelectronics.