Accurate surrogate models for the flat rolling process

Surrogate models, both polynomial and ANN-based (artificial neural networks), are developed to predict the rolling load in cold rolling of flat metals. An accurate but fast model was developed to serve as high-fidelity model for the training of the machine learning algorithms, allowing for large sampling sizes (up to 1000 samples) with different sampling methods, a number of eight input parameters, and various configurations of surrogate models. The ANN-based models have shown excellent predictive abilities provided that the training sampling is sufficiently large (more than 500 elements). In contrast, polynomial models converge much rapidly to their optimal accuracy (samplings of tens of elements) but their predictive ability is more limited, unless the order of the polynomials is increased. The latin hypercube sampling was more efficient than the random sampling in all cases.