International Society for Industrial Process Tomography

Estimation of Porous Material Parameters Using Ultrasound Tomography and Deep Learning

T. Lähivaara^1*, L. Kärkkäinen^2,3, J. M. J. Huttunen², and J. S. Hesthaven⁴

¹Department of Applied Physics, University of Eastern Finland, Kuopio, Finland

²Nokia Bell Labs, Espoo, Finland

³Department of Electrical Engineering and Automation, Aalto University, Espoo, Finland

⁴Computational Mathematics and Simulation Science, Ecole Polytechnique Fédérale de Lausanne,

Lausanne, Switzerland

^*Email: timo.lahivaara@uef.fi

ABSTRACT

We investigate the feasibility of simulation-driven machine learning to ultrasound tomography in order to estimate water-saturated porous material parameters.The problem consists of two main components: 1) the forward model that approximates wave propagation in coupled poroviscoelastic-viscoelastic-acoustic media from source to receivers and 2) the estimation problem where the prediction of the physical parameters are made using recorded data. As the forward solver, we use a discontinuous Galerkin method whereas deep convolutional neural networks are used to solve the parameter estimation problem. In the numerical experiments, we estimate the material porosity and tortuosity while the remaining parameters, assumed to be of less interest, are successfully marginalized. Results confirm the feasibility and accuracy of this approach.

Keywords Ultrasound tomography, Deep convolutional neural networks, Discontinuous Galerkin method, Porous materials

Industrial Application General