The recycling and disposal of polymeric materials have long represented a critical challenge for scientific research aimed at achieving the sustainable development goals of modern industry and society. Polymers are used in a wide range of applications, from everyday packaging such as polyethylene terephthalate (PET) to advanced engineering sectors, including automotive, electronics, and aerospace, where high-performance composites like carbon fiber-reinforced polymers (CFRPs) are employed. As a result, the global volume of plastic waste remains extremely high, posing major environmental and technological challenges.
Among emerging recycling strategies, solvolysis in near-critical water has shown great promise for the recycling of materials such as PET and epoxy-based composites. However, these processes are governed by complex thermal and chemical mechanisms occurring across multiple length and time scales, which remain poorly understood at the molecular level.
To address this, the project aims to develop a machine learning potential for molecular dynamics simulations of PET solvolysis under sub and supercritical water conditions. Training data will be generated using ab initio molecular dynamics, supported by enhanced sampling techniques. The development workflow will incorporate active learning strategies, to minimize training costs while maximizing accuracy and transferability. Model development will rely on advanced deep learning techniques, specifically the MACE framework and the Neural Equivariant Potential (NEP) model implemented in GPUMD, which will be employed and compared as two state-of-the-art approaches for training a reactive potential capable of capturing the relevant thermochemical behavior.
Machine learning models, particularly those based on deep learning, have recently emerged as powerful tools in materials science, providing a way to bridge the gap between the accuracy of ab initio methods such as density functional theory (DFT) and the computational scalability of classical molecular dynamics (MD), thus enabling realistic and efficient simulations of chemically reactive systems.
The requested HPC resources are essential to carry out ab initio MD simulations, support active learning cycles, train deep-learning potentials, and perform large-scale MD simulations to investigate reaction pathways and transport phenomena involved in supercritical solvolysis.
This work is part of the EU-funded EURECOMP project (Horizon Europe, https://eurecomp.eu/home), which promotes innovative and sustainable strategies for the recycling and reuse of advanced composite materials.
Matteo Fasano, Politecnico di Torino, Italy