Neural Network Surrogates for Uncertainty Calibration in Soft Tissue Mechanics

Soft biological tissues show complex mechanical behavior, including inelastic, time and history dependent characteristics. Advanced constitutive models can capture these responses, but identifying model parameters is challenging in general, and even more if experimental data are limited: Different parameter combinations may describe the available data equally well, however, their predictedresponse under other unseen loading conditions can be highly different.

This project aims to develop methods for the efficient calibrations of established complex material models for soft biological tissues including the quantification of uncertainty, which is caused by the limited database for calibration. A primary goal is to enable such analyses at relatively low computational cost so that the technique can be used to determine “likely” sets of material parameters rather than a unique combination which reproduces the calibration data.

To this end, we develop surrogate modeling techniques based on artificial neural networks and use them for uncertainty analysis based on Monte-Carlo methods. We thereby combine the proven reliability of established constitutive models with the effectiveness of modern machine learning techniques.