A machine learning route to identify the shape of metal nanoparticles

The global structure of nanomaterials can directly influence their physical and chemical properties. Among them catalytic properties depend on the size and shape, as well as on the local structure and environment of catalytic sites. Given the structural versatility of metal NPs as a function of synthesis conditions, the experimental shape recognition of nanoparticles is key. On the way to an automatic classification of the metal core of inorganic NPs from experimental data, a supervised machine learning model trained from numerical simulations was designed.

Four different supervised artificial neural networks were trained, tested, and submitted to a challenging dataset that were 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 that was tested allowed to classify nanoparticles based on its discretized RDF profile and its first derivative. 100% accuracy is reached on the test phase, and up to 70% accuracy is obtained on the challenging dataset. This last dataset was made of compounds that have global shapes significantly different from the training set. However, some nonobvious structural patterns make them related to the eleven classes learned by the ANNs. After some adaptation, this strategy may open the way to the recognition of nanoparticles based on experimental neutron or X-ray diffraction data.