Dynamic Properties of Elastic Media Obtained with Self-Trained Convolutional Neural Networks

This research will derive new knowledge related to the non-destructive evaluation of all the mechanical parameters that fully describe the elastic behavior of complex media such as biological tissue and designer materials with carefully engineered microstructure. Previous methods for non-destructive evaluation often rely on limiting assumptions on the nature of the probed material (for example, direction-independent properties), require large samples for evaluation, and/or provide only a small subset of material parameters.

For example, accurately evaluating all the mechanical parameters of biological tissue is important because it informs on tissue changes due to pathological processes. However, state of the art methods used to probe these parameters (e.g. elastography) may provide only a small set of mechanical parameters, which may delay the detection of tissue pathologies.

This project will create new methods to extract the complete set of material properties from scattered ultrasound pulses processed with convolutional neural networks trained in an unsupervised manner. The results of this research will be leveraged by numerous fields ranging from infrastructure integrity evaluation to non-invasive medical diagnostics and thus will benefit the society at multiple levels.

This research will also provide the opportunity to develop an online wave dynamics lab accessed and controlled remotely, which will increase the participation of economically disadvantaged students and underrepresented minorities to cutting-edge experimental science.

Analytical methods currently used to extract the dynamic mechanical properties of matter from scattered mechanical waves have proven insufficient when several dozen parameters are unknown. This effort will derive new knowledge and tools necessary to extract all the unknown parameters of complex media with anisotropic stiffness, mass density, and Willis parameter tensors.

The central hypothesis is that convolutional neural networks can learn the very complex mapping scattered-fields-to-constitutive-parameters from numerical simulations. This simulation-based approach will lead to an unsupervised self-training process that contrasts with most machine learning applications requiring expensive training data sets typically labeled by humans and obtained in long measurement sessions.

This hypothesis will be verified by fabricating and extracting the material parameters of elastic metamaterials designed to have anisotropic stiffness, mass density, and Willis parameter tensors. This research will consider material property extraction from both near- and far-field measurements.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.