Extraction of physical properties of solids and liquids through deep learning from PVS and LPVS experiment

Abstract

PVS (Pendulum-type Viscoelastic Spectroscopy) and LPVS (Liquid Pendulum-type Viscoelastic Spectroscopy) are experimental methods to characterize respectively solids and liquids for their time-dependent and frequency-dependent material properties such as Young’s modulus shear modulus and viscosity In order to better solve the inverse problems to obtain material properties from experimentally measured raw data an artificial intelligence technique based on deep neural networks (DNN) has been developed Numerically calculated complex Young’s modulus shear modulus and viscosity from the finite element method in conjunction with experimental data are used to generate sufficient labeled data for DNN training Sufficiently reasonable convergence results are obtained from the training Improvements may be achievable through better DNN architecture design and training with more datasets It is demonstrated that DNN methodologies are powerful to solve inverse problems when other inverse schemes are not efficient For PVS inverse problem properties of PMMA specimens are predicted by DNN errors within 5 GPa in Young’s modulus E 0 1 in Poisson’s ratio and 0 01 in loss tangent has been reached so far For LPVS inverse problem mean particle size (D50) of ZrO2/MEK suspension fluid is predicted error within 5nm has been achieved so far

Date of Award	2020
Original language	English
Supervisor	Yun-Che Wang (Supervisor)