Inverse Design of Silver Nanoparticles Using Multi-Target Machine Learning

Inverse design, where we can prescribe the structure based on a desired property, is a singular ambition of data-driven material design, and if successful will finally enable market-pull innovation as opposed to technology-push. This is extremely challenging however, since there could be numerous combinations of materials characteristics that present the same properties and predicting a structure/property relationship does not distinguish between them. It is even more of a challenge in nanomaterials design since the design space is even greater, and inverse property/structure relationships will typically need to encompass multi-functionality. Optimization based on hypothetical databases predicted using machine learning has shown promise in recent years, but suffers from a high computational cost, lack of specificity, and no guarantees that the optimal material has been found. There are simply too many unknown variables (structural characteristics) and not enough known variables (properties) for optimization to be reliable. In this presentation we will describe an alternative approach to inverse design that overcomes each of these limitations. By drawing on the multi-functionality of nanomaterials and using multi-target machine learning methods, we develop a workflow that is fast, easy to use, and predicts the characteristics of a single nanoparticle that simultaneously meets a set of performance criteria, with a fault tolerance. The method focusses the outcome on the most important characteristics in an entirely data-driven way, and with comparable accuracy and generalizability as traditional forward structure/property machine learning predictions. We have demonstrated the new inverse design workflow on silver nanoparticles, to design samples that will simultaneously deliver prescribed values of the ionization potential, electron affinity, electronic band gap, energy of the Fermi level and the formation energy.