Machine learning exploration of the chemical space of single-ion magnets

Abstract

Single-molecule magnets (SMMs) are molecules displaying slow magnetic relaxation and magnetic hysteresis with potential applications in spintronics, high-density storage units, and quantum computing. However, SMMs can only retain their magnetic hysteresis at very low temperature, with the current record holder at 80 K. To improve and discover new SMMs with long relaxation times, machine learning models and data-driven techniques can be applied to navigate the vast chemical space and design new molecules with the desired magnetic properties. In this work, we present a high-throughput workflow capable of assembling structures from predefined ligands and metal ion(s) and predicting the magnetic properties (e.g., D tensors) for each structure. To do this, we designed a workflow to sample structures from the Single Ion Magnet Data Visualization (SIMDAVIS) dataset, the Crystallographic Open Database (COD), and the Cambridge Structural Database (CSD), perform quantum chemistry calculations, and obtain the corresponding magnetic properties. The entire workflow is fully automatized, being able to seamlessly filter any databases for the desired single ion molecules, generate input files for the necessary quantum calculations, and analyze the outputs of the calculations. Our approach is demonstrated to be efficient, with minimum human intervention, and easily generalizable. In addition to building a dataset, it provides data-driven guidelines for future synthesis of SMMs, instead of relying on intuition for molecular design. Then, to make use of the data obtained, a robust machine learning model able to predict magnetic properties efficiently is required. We explored a number of different machine learning models, ranging from the linear regression model to neural networks (NNs), to predict tensorial magnetic properties. We applied our model to various datasets to demonstrate the generalizability of the model. Finally, we applied machine learning to significantly speed up the spin relaxation simulation of a single molecule magnet and study its magnetic relaxation.