Capturing variability in material property predictions for plastics recycling via machine learning

Plastic mechanical recycling is the conventional technological step towards circularity. In such aspects, complex mixtures of polyolefin blends are often fed into mechanical recycling systems, resulting in moulded products with uncertain quality. To add to the difficulty of heterogeneous feedstocks, the testing of mechanical properties for plastic products often results in stochastic measurements, making connections from material prediction to systems understanding challenging. This research is aimed at providing a framework capable of generalizing stochastic plastic recycling knowledge via interval-based machine learning for the prediction of properties formulation for unrecycled plastics. The framework is made up of two components: a regressor for point estimation and an interval predictor for generating prediction intervals. We compare several competing methods for each of these components through empirical evaluation on a real-world dataset. The results demonstrate the usefulness of interval-based machine learning in the application of stochastic engineering problems such as plastic mechanical recycling, highlighting such approaches towards better model interpretation and (un)certainty prediction regions. • Investigated the use of machine learning for modelling the variability in material properties in the context of plastics recycling. • Proposed a framework for generating prediction intervals for plastic blends. • Performed exploratory data analysis to gain insight into our obtained dataset. • Performed a multi-metric quantitative evaluation of various competing components of our framework to discover best performing configurations. • Demonstrated that our proposed framework is capable of capturing the variability in material properties through prediction intervals.