Artificial neural network analysis of the catalytic efficiency of platinum nanoparticles

Even using high throughput methods, data-driven predictions of nanomaterials properties from first principles simulations can be impractical. In this work, machine learning models are developed to map the catalytic efficiency of Pt nanocrystals to structural features, such as nanoparticle diameter, surface area, sphericity, facet configuration and type of surface defects, using a theoretically derived big data set of over three hundred thousand nanoparticles. Artificial Neural Networks (ANNs) were calibrated with 50% of a data set including structural features of symmetric Pt nanoparticles; and catalytic activity, selectivity and thermodynamic stability. Surface response analysis was applied to two-inputs ANNs with squared correlation coefficient > 0.9, yielding a region of optimal catalytic efficiency for the less spherical nanocatalysts and {110} facets lower than 20%. Binary decision tree models reveal the optimal three-property combinations for high catalytic efficiency. In addition, ANN models built for non-symmetric nanoparticles predict the catalytic efficiency and stability with accuracy >0.93. In general, we show the combination of machine learning models can rapidly estimate functional properties of hypothetical nanomaterials at a resolution that is inaccessible to both computation and experimental methods, as well as identifying principles or rules that could guide rational nanomaterial design in the near future.