Data-Driven Design of Classes of Ruthenium Nanoparticles Using Multitarget Bayesian Inference

Producing perfectly regulated nanoparticle samples on a large scale is challenging and costly for manufacturers, so the ability to define and reproduce classes of nanoparticles with similar characteristics is attractive. However, developing structure/class or process/structure/class relationships is not straightforward. In this study we propose a machine learning pipeline of grouping nanoparticles based on their similarity in a high-dimensional feature space via clustering, predicting the nanoparticle classes from their structural features via classification, and identifying the relevant features that should be tuned to produce a specific class via causal inference. Using a simulated ruthenium nanoparticles data set as an exemplar, a support vector machine trained on 22 structural features managed to achieve highly accurate classification of ruthenium nanoparticles into ordered crystalline, polycrystalline, and disordered noncrystalline nanoparticles with virtually no overfitting and underfitting and high precision and recall. A Bayesian network with domain knowledge incorporated via interactive learning was trained using a hill climbing algorithm to confirm which features are causing the classes, as opposed to being just correlated to them.