Abstract Scope |
Bandgap is a crucial property of crystals, especially in photovoltaics (PVs) for which the bandgap largely determines the solar cell efficiency. Exploring large chemical spaces for PV materials is infeasible because conventional methods for determining the bandgap involve expensive quantum mechanical calculations. Recent efforts in computational materials science have explored using machine learning techniques to predict properties of materials. While these approaches have shown great promise, they have largely been limited to simple molecules with large training sets. We study a dataset of 62K molecular crystals and use the structure information to predict the molecular gap as a proxy for bandgap. We train a deep-learning model to achieve a mean absolute error of 0.17eV and develop an active learning framework that, given a pool of candidate structures, iteratively proposes the molecules that will provide the greatest reduction in prediction error, which requires fewer calculations to reach the same accuracy. |