Exchange coupling between unpaired electrons plays a crucial role in determining the magnetic properties of molecules, e.g., dinuclear copper(II) complexes, deciding their usability as molecular magnets in fields such as molecular spintronics or quantum technologies [1].
Problem-tailored machine learning models based on Gaussian process regression have proven capable of predicting the strength of the exchange spin coupling, expressed via the Heisenberg exchange coupling constant J and calculated using an efficient Green’s-function approach, within subsets of dinuclear copper complexes. Training the machine learning models requires only a fraction of the computational cost and produces accuracies close to the reference calculations obtained by utilizing density-functional theory (DFT) [2,3].
However, more complex datasets (which better represent the relevant chemical space of dinuclear copper complexes) reveal significant shortcomings of previous approaches regarding machine learning algorithm and descriptor choice. Examining the performance of machine learning models trained on increasingly complex datasets pinpoints the limitations of previously used low dimensional, problem-tailored descriptors and allows for systematic improvements while retaining low dimensionality and feature interpretability.
In particular, encoding the influence of outer ligands remains challenging due to their different structural as well as electronic effects on the coupling constant [4]. Simple solutions such as one-hot encoding (including selected element or ligand specifications) leads to decent performance improvements, highlighting the importance of proper feature selection, but does not adequately translate the unique properties of the many possible ligands.
Eduard Hahn