Virtual Experimentations by Deep learning on Tangible Materials
Abstract Artificial intelligence is an emerging frontier in material science to discover new materials with targeted properties by an artificial neural network (ANN) constructed from existing structure-property databases. This approach has not been applicable to tangible materials, such as plastic composites, fabrics, and rubbers, because the complexities of their structures cannot be defined. Here we propose a deep learning computational framework that can implement “virtual” experiments on tangible materials (carbon nanotube (CNT) films) where structural representations (scanning electron microscope images at 4 levels of magnifications (x2k, x20k, x50k, x100k)) of the processed material (dispersing and filtering) were created by multiple generative adversarial networks from which an ANN predicted multiple properties (electrical conductivity and specific surface area). 1865 virtual experiments were finished within an hour, a task that would take years for real experiments. The accumulated data can be used as a versatile database for material science, in analogous to databases of molecules and solids used in cheminformatics, as exemplified by investigations of the correlation between the electrical conductivity and specific surface area, wall number phase diagrams, the most economical mixture of CNTs at specific property, and inversely designed CNT supercapacitors.