MCA: Physics-Informed Machine Learning from Acoustic Data

This Mid-Career Advancement (MCA) grant will fund research and training that enables effective use of acoustic sensors to estimate material properties, diagnose state, or predict failure in engineered devices and structures, as well as for seismic monitoring of mines, carbon sequestration sites, and geothermal energy reservoirs, thereby promoting the progress of science and advancing the national prosperity and welfare. Acoustic sensors range from miniature transducers for medical diagnostics and structural health monitoring to seismometers for recording large-scale ground motion. In current practice, the time-varying signals from such sensors are reduced from thousands of data points to a few hand-crafted features. This underutilization of data results in poor feature resolution and an inability to accurately diagnose the evolving state of materials or predict imminent failures. This project will overcome such shortcomings by building an analysis and modeling framework that accounts for the full acoustic signal waveform and is informed by knowledge of the underlying physics, thereby achieving improved accuracy, generalizability, and interpretability of its predictions. This framework offers an unconventional approach to nondestructive defect detection in aerospace, automotive, infrastructure, pipelines, and energy industries, as well as to the longstanding challenge of earthquake prediction, and may translate into improved classification of abnormalities in medical ultrasound imaging. An integral component of the training plan is the development of a new project-based graduate-level course bridging acoustics and machine learning and with course materials made available online.

This research aims to make fundamental contributions to the integration of machine-learning techniques with domain-specific knowledge about the elastodynamic response of complex materials systems in a data analysis framework that extracts an information-rich set of critical features from acoustic data. It achieves this goal by constructing physics-informed deep learning models to predict shear failure using ultrasonic data from laboratory experiments on rocks. Such models are expected to minimize the risk of overfitting, thereby making them adaptable also to other datasets, and to be more explainable, permitting new understanding of wave propagation and interactions in complex media, for example, with heterogeneities and discontinuities. First, multi-headed models that incorporate dimensional reduction techniques and multi-task learning are trained to identify the most informative acoustic features in a limited dataset. Next, models are constructed that also respect coupled friction and 3D wave transmission laws. Finally, model interpretability, robustness, and generalizability are evaluated against a larger set of experimental conditions within and outside the bounds of the training data.

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.