(611d) Towards an Open-Source Implementation of Spatially-Resolved Molecular Fingerprinting Methods in Machine Learning-Based Predictions of Material Properties

The fields of artificial intelligence (AI) and machine learning (ML) have experienced tremendous growth and success over the past few decades. This has led to the development of innovative and powerful solutions to a host of problems in various fields including robotics, medicine, manufacturing, etc. AI and ML also show great promise for applications in materials design and characterization. One such possibility is to help uncover characteristic structure-property relationships in various classes of materials. An important roadblock towards ML-based direct mapping between structures and properties is the availability of numerical representations that can capture various structural signatures while providing spatial resolution and being invariant to operations such as geometric rotations and translation. Cartesian coordinates traditionally used to represent molecular structures suffer from a lack of uniqueness and variance under such property-preserving geometric operations. Over the past decade or so, progress has been made in developing alternative representations, called “fingerprints,” that are suitable for ML applications. In this work, we review and compare fingerprints that provide spatial resolution and propose a new classification scheme to better grasp their applicability and limitations. Moreover, we introduce a new open-source software package called SEING, designed and implemented for easy computation of spatially resolved molecular fingerprints. We finish by showing example applications for force prediction in crystals such as Si, Al and W with a generalization accuracy on the order of 0.1 meV Å^−1.