Machine learning for viscoelastic constitutive model identification and parameterisation using Large Amplitude Oscillatory Shear (2312.13793v1)

Abstract: Identification and parameterisation of constitutive models can be a challenging task in rheology. We investigate the use of Random Forest (RF) regression to estimate viscoelastic constitutive model parameters using Large Amplitude Oscillatory Shear (LAOS) data. Specifically, we deploy the RF to predict constitutive model parameters using the spectra of Chebyshev coefficients pertaining to the stress-strain-strain rate Lissajous curves. As far as we know, this is the first time Machine Learning (ML) has been utilised for this predictive task. We test three constitutive models: the linear and exponential Phan-Thien-Tanner (PTT) models, and the RoLiE-Poly model. For both PTT models, the RF estimator predicts the model parameters from the Chebyshev spectra with high accuracy. For the RoLiE-Poly model, the RF estimator demonstrates lower accuracy in model validation, however the predicted parameters still accurately capture the rheological behaviour in both oscillatory shear and oscillatory extensional flow. This is because multiple sets of RoLiE-Poly model parameters can yield very similar rheological responses and indicates practical identifiability issues under the investigated conditions. This does not represent an inadequacy of the framework, but rather a fundamental challenge in estimating model parameters using rheometric data. Overall, the study highlights the strong potential ML has to offer for selecting and parametrising constitutive models based on rheometry data. Our results can help in allowing for widespread robust, modelling of viscoelastic fluid flows.