Dr. Bergstrom, I have great respect for your work. I myself have learned about polymer mechanics from your publications. My comment is to kindly defend, to some extent, models based on artificial neural networks: I believe the comparison you present does not truly explain the power of these models. While it's true that an ANN like the one you demonstrate is too computationally costly to model a material using only a single variable as input, the reality is that these models are useful when there are no constitutive models for variables beyond deformation (even with rare cases like bacterial attack on a material). We all know that modeling materials can be time-consuming, and ANNs help solve problems quickly; although this doesn't mean one should not eventually seek more succinct and explainable constitutive models than these. In this sense, I think it would be fair to compare an ANN with advanced constitutive models that take more than one variable as input. Regards!

Another interesting approach could be to use Bayesian machine learning with Markov chain Monte Carlo (MCMC) sampling to obtain posterior distributions for for all parameters from which credible intervals for the parameters as well as stress could be obtained. This would show how much error would be in the calibration; one could assess if the model is correct and/or if the data set is suitable. The downside is that the MCMC process would take longer to compute than for a typical calibration.