Sustainable Valorization of Polyethylene Waste: Proof-of-Concept for an Integrated Simulation and Machine Learning Framework for Olefin Recovery

Abstract

The valorization of plastic waste into olefin feedstocks is a critical enabler of the circular economy, yet machine-learning surrogates for downstream olefin recovery remain prone to training on nonphysical simulation artifacts, and previous studies have relied on small datasets without rigorous cross-validation. This work addresses this gap by developing a validated artificial neural network (ANN) surrogate model for polyethylene plastic waste thermal cracking followed by distillation in Aspen HYSYS. In total, thirty-one operating conditions were selected using Latin Hypercube Sampling; seven cases failed to meet physics-based criteria (mass balance, recovery rate, energy utilization ratio), resulting in 24 successful cases. For the development of an ANN (input parameters: eight; hidden layer neurons: ten; output: eight) that employs L2-regularized Levenberg–Marquardt optimization and 30-repetition leave-one-out validation, the performance metrics of the surrogated distillate flow rate (R2 = 0.828 [0.135]), reboiler heat duty (R2 = 0.852 [0.065]), and 1-butene purity (R2 = 0.814 [0.093]) met screening-level criteria, while the condenser temperature (R2 = 0.419 [0.196]) required hybrid Peng–Robinson flash calculations due to near-critical VLE sensitivity. Sensitivity analysis with respect to the rigorous dataset shows that, within the investigated ranges, feed composition is a more important variable in olefin recovery than reflux rate or the number of theoretical trays. The estimated annual reduction in net GHG emissions in terms of gate-to-gate analysis is 207,000–337,000 tonnes CO2-eq per 131,600 tonnes PE processed (incineration counterfactual) or 165,600–273,700 tonnes (landfill counterfactual), primarily through fossil naphtha displacement (∼106,000 tonnes/year); no landfill-methane credit is claimed. This study demonstrates a transferable, physics-informed machine learning framework for screening-level design of plastic waste valorization processes.