hide
Free keywords:
-
Abstract:
Machine learning (ML) has been widely used for chemical property prediction. In the context of molecular sciences, the prediction of simple electronic properties (e.g. the atomization energy) has reached extremely high accuracy for small and rigid molecules. However, the accurate prediction of more complex electronic properties, which are important targets for molecular/materials design, still poses a challenge. This is especially true for highly flexible molecules, due to their vast conformational variations and the intricacies of the underlying structure-property relationships. Examples of such properties are orbital energy levels, electronic couplings and molecular reorganization energies, all of which display markedly different properties than more common chemical ML targets like atomization energies. There is thus significant demand for methodological developments in chemical ML, in order to leverage its potential for molecular and materials design, for example in the context of organic semiconductors (OSCs).
In this work, we specifically focus on two molecular properties which are known to influence the performance of OSCs, namely the energy of the highest occupied molecular orbital (HOMO) as well as the internal reorganization energy (λ). First, we explore strategies to improve the performance of ML models for λ prediction by using semi-empirical methods for conformer sampling and as a baseline. The obtained models are then evaluated in a diverse chemical space to discover novel low-λ structures. This also enables the discovery of general chemical design rules by substructure searching among favourable candidates. Second, we consider the suitability of state-of-the-art ML models for predicting the HOMO energy. Being an intensive and sometimes localized property, we show that the common pooling strategies in atomistic neural networks are ill-suited for this target. To overcome this issue, we propose a series of physically motivated pooling functions. Among these, the novel orbital weighted average (OWA) approach is developed. OWA enables accurately predicting the orbital energies and distributions simultaneously. The underlying approach is also promising for other intensive properties such as excitation energies.
