Surfing Multiple Conformation-Property Landscapes via Machine Learning: Designing Single-Ion Magnetic Anisotropy

Abstract

The advent of computational statistical disciplines, such as machine learning, is leading to a paradigm shift in the way we conceive the design of new compounds. Today computational science does not only provide a sound understanding of experiments, but also can directly design the best compound for specific applications. This approach, known as reverse engineering, requires the construction of models able to efficiently predict continuous structure-property maps. Here we show that reverse engineering can be used to tune the magnetic properties of a single-ion molecular magnet in an automated intelligent fashion. We design a machine learning model to predict both the energy and magnetic properties as function of the chemical structure. Then, a particleswarm optimization algorithm is used to explore the conformational landscapes in the search for new molecular structures leading to an enhanced magnetic anisotropy. We find that a 5% change in one of the coordination angles leads to a ∼50% increase in the anisotropy. Our approach paves the way for a machine-learning-driven exploration of the chemical space of general classes of magnetic materials. Most importantly, it can be applied to any structure-property relation and offers an effective way to automatically generate new materials with target properties starting from the knowledge of previously synthesized ones.

Exchange Spin Coupling from Gaussian Process Regression

“…A possible solution to this problem are machine-learning (ML) approaches 31,32 , which have seen an enormous gain in popularity and development in the past few years in the chemistry, materials science and solid-state physics communities , and which may save orders of magnitude in computing time compared to DFT, and often comparable or even better accuracy. ML has proven successful for spin-dependent molecular properties, in particular for spin-state energies in spin crossover complexes 35,36,[54][55][56][57] , magnetic moments and magnetic anisotropy 58,59 , and also for properties closely related 60,61,[61][62][63][64][65][66][67][68] to exchange spin coupling such as charge transfer [69][70][71] and excitation energy transfer 71,72 . The capability of ML for exchange spin coupling has not been explored yet.…”

Exchange Spin Coupling from Gaussian Process Regression

“…For the spin state energetics in spincrossover compounds, revised two-dimensional autocorrelation functions combined with feature selection algorithms have proven as very successful, both with kernel ridge regression and with artificial neural networks 57 . For learning magnetic anisotropy in single-ion molecular magnets, bispectrum components combined with ridge regression were shown to be a good choice 59 . As a first step towards exploring the machine-learnability of J, we will focus on a selection of descriptors which have proven successful for a variety of properties and structures, and which are inexpensive to evaluate computationally: Smooth overlap of atomic positions (SOAP) 113 , many-body tensor representation (MBTR), and manybody interaction descriptors 114 (F 2B and F 3B ).…”

Machine Learning in Magnetic Materials

“…This model was then used to explore the molecular conformational landscapes in search of structures that maximize magnetic anisotropy. [ 77 ] Moreover, Lunghi and Sanvito established a method based on ML and electronic structure theory that makes the prediction of spin lifetime in realistic systems feasible. [ 78 ] Nelson and Sanvito used an ML approach to overcome the incapacity of typical high‐throughput electronic structure calculations to estimate the Curie point, the temperature where a ferromagnetic material loses its magnetic properties.…”