Learning from the Machine: Uncovering Sustainable Nanoparticle Design Rules

Machines consisting of bags of artificial neural networks (ANNs) have been constructed to connect nanoparticle features to the viability of a broad class of organisms upon exposure. The optimization of these machines is based on a relatively small data set. However, through consensus across a bag of ANNs, these machines predict at a level of confidence comparable to the experiment and perform better than chance. The mining of the machine across the feature space allows for the discovery of design rules for nanoparticles with increased viability. As such, we demonstrate the efficacy of inversion as an approach to learn from the machine in the context of designing sustainable nanoparticles. For example, we find that increased manganese content in lithium NiMnCo oxide nanoparticles is associated with greater viability, carbon dots reduce viability less than quantum dots, and gold nanoparticle coatings can significantly affect viability at high concentration.