2025 AIChE Annual Meeting

(588ax) Studying Depolymerization of Polyurethanes Using Reaction-Aware Deep-Learning Potentials

Authors

Siddarth Achar, University of Pittsburgh
Karl Johnson, University of Pittsburgh
Christopher E. Wilmer, University of Pittsburgh
Polyurethanes are one of the most widely used classes of polymers. The polymerization of polyurethanes (PUs) is a relatively simple process, facilitated by the formation of strong urethane bonds between constituent moieties. This strong urethane linkage enables a wide range of applications1—from rigid insulating materials to soft foams. However, the high stability of the urethane bond also makes it non-degradable, posing a significant environmental concern when it comes to disposal. This highlights the importance of chemical depolymerization as a recovery method to reduce waste accumulation and mitigate the environmental impact of this polymer.

Our work focuses on developing a framework for efficiently training reaction-aware deep learning potentials2 (DPs) that can be used to investigate various reactions and reaction pathways for PU depolymerization. We combine density functional theory (DFT) and machine learning to study these reaction pathways, since using DFT alone to evaluate new reactions is impractical due to its limitations in time and length scales. We employed the single-ended growing string method3 (SE-GSM), using our DPs as a calculator, to explore reaction pathways involved in polyurethane decomposition.

We began by training bootstrap DPs using data from density functional theory-molecular dynamics simulations. Next, we performed molecular dynamics active learning (MD-AL) with the DPs to enhance their performance. Unimolecular decomposition reactions and driving coordinates are enumerated using the Yet Another Reaction Program4. These driving commands were then used with the extended tight binding semi-empirical method5 to carry out SE-GSM calculations. We sampled intermediate structures, relabeled them using DFT, and retrained our DPs to generate gen 0 DPs. These gen 0 DPs were then used to perform DP-GSM calculations and sample additional intermediates. These intermediates were again relabeled using DFT and used to retrain the DPs, resulting in gen i (i = 1,2, 3... and so on) DPs. We also conducted self-consistent MD-AL cycles using the stable intermediates sampled from the DP-GSM calculations.

This enables exploration of the reactive space through automated generation of potential PU degradation reactions and application of SE-GSM. Through multiple reactive active learning and self-consistent MD-AL loops, our DPs have explored most of the relevant configurational and chemical space for PU degradation. The trained DPs can accurately predict reaction barriers for PU degradation via our active learning scheme. This lays the groundwork for future efforts to identify conditions that enable efficient depolymerization and recycling. With further development, this approach may also help improve the selectivity and efficiency of PU breakdown processes.

References:

  1. de Souza, F. M., et. al. American Chemical Society (2021)
  2. Wang, H., et. al. Computer Physics Communications (2018)
  3. Zimmerman, P. M., Journal of computational chemistry (2015)
  4. Zhao, Q., et al. Nature Computational Science (2021)
  5. Bannwarth, C., et. al. Wiley Interdisciplinary Reviews: Computational Molecular Science (2021)