Active Learning Approaches to Predict Intermolecular Noncovalent Interactions in Organic Semiconductors

The natures and strengths of intermolecular noncovalent interactions are critical to the formation of organic semiconductors (OSC) from their molecular or polymer-based building blocks. While there are several quantum-chemical approaches available to evaluate intermolecular noncovalent interactions, including symmetry-adapted perturbation theory (SAPT), these methods can be computationally expensive, especially for the types of large building blocks used in OSC. This computational cost limits the ability of machine-enabled searches of the OSC chemical and materials space. Machine learning (ML) models have emerged as powerful tools to provide fast predictions of a variety of molecular and material properties at a fraction of the computational cost when compared to quantum-chemical approaches, though they often require substantially large and labeled datasets, posing challenges when data is scarce or difficult to obtain. Here, we discuss an active learning approach that augments ML models trained to predict intermolecular noncovalent interactions by incorporating a prediction confidence. The active learning scheme is developed to enable the reduction of labeled data requirements by identifying regions in the chemical space where the model exhibits the most uncertainty. This combined approach demonstrates effective and fast determination of intermolecular noncovalent interactions in OSC.