PolyTransGAN enables neuro-symbolic inverse design of 4D-printable shape-memory polymers

Abstract

While the shape-memory polymer (SMP) inverse design is promising for 4D printing, it is hindered by the large chemical space and the lack of physics-aware generative models. We introduce a neuro-symbolic method, called PolyTransGAN, which combines coarse-grained molecular dynamics (CG-MD) and a conditional GAN with a physics-guided conditional generator. A Physics consistency loss, inspired by the Fox equation is directly incorporated into the training objective, ensuring thermodynamic plausibility when generating the sequence. The model is trained with a high-fidelity dataset of 50,000 MD-validated polymer topologies, called SMP-4D. PolyTransGAN achieves a 78% success rate for errors below 15% for both transition temperature (T_trans) and recovery rate (R_r), 100% chemical validity and 97% uniqueness. PolyTransGAN achieves a computational speedup of 1200× compared to NSGA-II and high property matching accuracy. The validity of the presented model is confirmed by the results of the extended validation with 50 LAMMPS simulations, which show a median relative error of 10.1%, and 95% of the predictions are within the limits of ± 20% of the target values, especially in the range of 30–60 °C, which is important for technical applications. The results show that the neuro-symbolic integration approach is a scalable and physics-informed method for the inverse design of 4D printing functional polymers.