Physical Simulator-Based Neural Networks for Real-Time Fouling Tomography

We propose a real-time method for estimating the spatial distribution of fouling within an enclosed metal structure by analyzing amplitude changes in ultrasonic Lamb waves with a probabilistic machine learning model. The model is trained exclusively on synthetic data without requiring any labeled empirical data, instead leveraging simple physical properties of ultrasonic waves. For synthetic fouling distributions, following a suitable prior distribution, such as a Gaussian Process (GP), we form learning targets using a physical forward model characterizing the exponential decay of the signal amplitude along propagation paths determined by the problem geometry. We then learn a neural network for amortizing the costly GP inference needed for the inverse problem. In addition to speeding up inference, our formulation allows easy use of richer prior knowledge, such as prior expectation of specific fouling shapes or spatial distributions, by embedding suitable training examples into the synthetic training data. We evaluate the characteristic properties of the method on simulated data and verify its performance in a real ultrasonic sensing setup. We demonstrate an accurate estimation of high-resolution fouling maps in milliseconds on a standard laptop based on just four sensors by aggregating information across multiple helical wave propagation paths, making significant progress in structural monitoring.