Recognition of the three-dimensional structure of small metal nanoparticles by a supervised artificial neural network

Catalytic characteristics of metal nanoparticles heavily depend on their global shapes and sizes as well as on the structure and environment of catalytic sites. On the computational chemistry side, calculations of thermodynamic and kinetic data involve a high calculation cost which can be significantly lowered by the use of a trained machine learning model. This paper outlines a preliminary approach that aims at classifying the shape of the metal core of nanoparticles. Four different supervised artificial neural networks were trained, tested and submitted to a challenging dataset. They are based on two different structural descriptors, Coulomb matrices and radial distribution functions (RDFs). Each model is trained with hundreds of 3D models of nanoparticles that belong to eleven structural classes. The best model classifies a NP according to its discretized RDF profile and its first derivative. 100% accuracy is reached on the test stage, and up to 70% accuracy is obtained on the challenging dataset. It is mainly made of compounds that have global shapes significantly different from the training set. But some nonobvious structural patterns make then related to the eleven classes learned by the ANNs. Such strategy could easily be adapted to the recognition of NPs based on experimental neutron or X-ray diffraction data.