Deep learning for inverse material characterization

This paper presents an approach for computationally efficient inverse material characterization using Physics-Informed Neural Networks (PINNs) based on partial-field response measurements. PINNs reconstruct the full spatial distribution of the system’s response from the measured portion of the response field and estimate the spatial distribution of unknown material properties. The primary computational expense in this approach is the one-time generation of potential responses for the PINNs, resulting in significant computational efficiency. Furthermore, this study utilizes PINNs to train a model based on the underlying physics described by differential equations, and to quantify aleatoric uncertainty arising from noisy data. We demonstrate several one-dimensional and two-dimensional examples where the elastic modulus distribution is characterized based on static partial-field displacement response measurements. The inversion procedure efficiently provides accurate estimates of material property distributions, showcasing the potential of PINNs in practical applications.