Investigation of transition metal complex representations for machine learning structure-property relationships

Molecular magnets have potential applications in quantum computing, spintronics, and sensor development. These molecules display spin anisotropy below their characteristic blocking temperature. Contenders for single molecular magnets are monometallic transition metal complexes. Modeling of these complexes demand high computational cost and is difficult due to strong coupling effects. We investigate the performance of crystal graph neural networks (CGNN) for the prediction of properties using a dataset containing nearly 87,000 transition metal complexes. These properties have been calculated using the TPSSh-D3BJ exchange-correlation functional. Here, we see if the CGNN can predict the HOMO/LUMO gap, metal ion charge, and a variety of other computed energies. We then compare the model performance of the CGNN against neural networks trained with structural descriptor representations, such as the smooth overlap of atomic positions (SOAP). A completed model can be used to filter complexes in a high throughput screening. This work provides the first steps in the development of a machine-learning model for the property prediction of transition metal complexes for single molecular magnet applications.