The prediction of defect harmfullness in continuum mechanics requires the solution of partial differential equations with specified boundary conditions. In this work, fast reduced-order models are developed to understand defect harmfulness in recycled polymers undergoing a stretch blow moulding process. Polymers underconsideration are PET. Such a process is used in bottle-to-bottle recycling [1]. The space of all possible defects is defined using a non-parametric, data-driven approach that takes into account defects seen using an infrared camera. This space is very high dimensional.

The aim of this work is to develop a non-linear dimensionality reduction approach as proposed in [2] and [3], by using a scientific machine learning. Numerical results show that the proposed reduced order models have a large validity domain in the parameter space related to the mechanical behaviour of recycled polymers. They also show computational speed-ups of around 50 compared to conventional finite element predictions. Accelerated defect damage predictionscan be used to predict potential process quality degradation.