Non-destructive Hardness Prediction via Deep Learning Image Regression Models

Over generations of evolution, nature has developed incredibly complex material structures fit to their environments. Using these time-tested designs as a base, human-engineered bioinspired structures are a promising avenue for pushing the boundary of possible materials configurations. However, how to best navigate the diversity of bioinspired structures to attain desired properties of interest remains an open question. Here we focus on one of the hardest biological tissues found in animals, tooth enamel, to examine the relationship between structure and property. While the typical indentation methods of measuring material hardness are time consuming and destructive to the sample, we propose artificial intelligence models will allow us to predict mechanical properties directly. We train an image regression neural network on known hardness indentation data as a non-destructive surrogate model that can predict the hardness of pristine structures much quicker than experiment. Subsequent visualization tools such as gradient ascent and saliency maps allow us to identify which regions of structure contribute the most to their hardness. In doing so, this approach not only demonstrates a method for improved material characterization, but also assists in understanding the connection between structure and property for engineering design problems.