Genetic algorithms and graph neural network applications for molecular magnetism

Abstract

Molecular magnets have been under active investigation for more than three decades due to their capability of storing magnetic information on a molecular level. However, their applications remain limited to this day, due to spin-phonon relaxation processes destroying this information. Calculating the full relaxation rate is an expensive task, and therefore, the discovery of high relaxation time molecular magnets is a slow process. However, it has been linked to the magnetic anisotropy of the compounds in question, allowing this property to be used as a proxy to speed up this process. We implement two methods that, while having found widespread application inside and outside of chemistry, have not been used in the discovery of molecular magnets: In a first step, we develop two flavors of genetic algorithms that are capable of discovering new compounds on their own using on-the-fly ab initio methods. The chemical space investigated can either be spanned by supplying a list of ligands that are assembled into a coordination compound or a set of rules allowing the algorithm to develop new ligands on its own. The resulting algorithms are benchmarked against a previously computed dataset, proving their speedup over random search methods, and then subsequently applied to Co(II) and Dy(III) compounds, discovering new and record-breaking compounds. In a second step, we develop different kinds of graph neural networks, capable of predicting the Kramers energy levels in Dy(III) compounds responsible for their magnetic behavior. These models are trained on a dataset specifically designed for this task, which also provides the largest dataset of Dy(III) compounds to our knowledge. The different models are fed different levels of complexity in input data, and can, in addition to the invariant Kramers energies, also predict the spin Hamiltonian itself, allowing for the computation of other magnetic properties as well. In conclusion, this work presents two novel algorithms that, to our knowledge, have not been applied to molecular magnets before, with the capabilities to speed up calculations significantly or reduce the amount of calculations required by sampling the chemical space more efficiently.