Explainable machine learning insights into molecular clusters nuclearity via pair distribution function

The pair distribution function (PDF) is a powerful tool for structural investigation of materials, extending the analysis of scattering data beyond conventional crystallography. While the PDF encodes the atomic arrangement, an initial model is often required to gain detailed insight into a material’s structure. The number of heavy atoms (nuclearity) is an important feature bridging coordination and nanocluster chemistry. This study describes the development of a convolutional neural network that achieved a mean nuclearity prediction accuracy of 86% after training on calculated PDFs of 645 crystal structures of lanthanide compounds. The trained network, when applied to experimental PDFs, successfully identified the nuclearity of compounds and demonstrated an emergent generalization ability across the simulation-to-experiment domain gap. The study widens the perspective on extracting information from the PDF and opens prospects for using machine learning tools for a wider scope of tasks in chemistry and materials science.