ABSTRACT
Many possible molecular crystal forms need very accurate energy for quantum mechanical prediction. Even though hundreds or even thousands of structures may exist in the low energy region of interest, the computational costs associated with calculating their energies using high performance density functional theory (DFT) are exorbitant. Here we use machine learning to reassess the energies underlying crystalline structure indicated by a cheap protective shield in order to foresee the need for time-consuming & costly hybrid modular DFT (PBE0) computations. To fill the gap seen between protective shield and the PBE0 energies, this technique makes use of a more advanced autoregressive Gaussian process (IAGP) and calculations that are more efficient using DFT (PBE). Given the probabilistic nature of our model, we explore how these variations in anticipated energies affect our ability to rank crystal structures according to their energetic merit.
