Virtual experimentations by deep learning on tangible materials

Artificial intelligence relying on structure-property databases is an emerging powerful tool to discover new materials with targeted properties. However, this approach cannot be easily applied to tangible structures, such as plastic composites and fabrics, because of their high structural complexity. Here, we propose a deep learning computational framework that can implement virtual experiments on tangible structures. Structural representations of complex carbon nanotube films were conducted by multiple generative adversarial networks of scanning electron microscope images at four levels of magnifications, enabling a deep learning prediction of multiple properties such as electrical conductivity and surface area. 1716 virtual experiments were completed within an hour, a task that would take years for real experiments. The data can be used as a versatile database for material science, in analogy to databases of molecules and solids used in cheminformatics. Useful examples are the investigation of correlations between electrical conductivity, specific surface area, wall number phase diagrams, economic performance, and inversely designed supercapacitors. Artificial intelligence may significantly accelerate the discovery of new materials but is not easily applicable to non-periodic structures. Here, a deep learning framework is proposed to predict properties of tangible carbon nanotubes by generating virtual structures at different scales and compositions.