Deep learning virtual indenter maps nanoscale hardness rapidly and non-destructively, revealing mechanism and enhancing bioinspired design

Over evolution, organisms develop complex material structures fit to their environments. Based on these time-tested designs, hu-man-engineered bioinspired structures offer exciting possible ma-terials configurations. However, navigating diverse structure spaces for attaining desired properties remains non-trivial. We focus on the hardest biological tissue in humans, tooth enamel, to examine the structure-property relationship. While typical hardness measure-ments are time consuming and destructive, we propose that artifi-cial intelligence models can predict properties directly and enable high-throughput, non-destructive characterization. We train a deep image regression neural network as a surrogate model and visualize with gradient ascent and saliency maps to identify struc-tural features contributing most to hardness. This model demon-strates improved spatial resolution and sensitivity compared with experimental hardness maps. Using this rapid hardness testing model, a generative adversarial model, and a genetic algorithm that operates in latent space, allows for guided materials design, yielding proposed designs for bioinspired structures with precisely controlled hardness.