Abstract

Much has been said about the ability of machine learning to reduce the computational cost of quantum mechanical calculations – by closely approximating the approximation level of reference data (within a domain of applicability close to the data manifold). In order to obtain the highest possible accuracy the SOTA prediction methods are all structure based – SOAP , SchNet (DTNN), MBTR , CGCNN/MegNet.

However for applications within materials discovery we often start without knowledge of the crystal structure and so new approaches are needed if we want to use machine learning to accelerate such workflows.

I will briefly summarise what has been done in this structure-free domain and then introduce a new end-to-end model that addresses some of the short comings. https://arxiv.org/abs/1910.00617