The long-time scale behavior of hydrogels is still a core problem in material design, especially due to the limitation of computational cost and accessibility time scale of molecular dynamics (MD) simulations. So, this study offered a combined model, which integrates both MD and artificial intelligence (AI) models to truly predict the dynamics of structural and transport characteristics of thiol-norbornene click cross-linked carboxymethyl cellulose (CMC) hydrogels. All-atom MD portrays a consistent network of hydrogen bonds, a measurable swelling ratio, and restricted polymer mobility, and these are indicators of a strong gel. To extrapolate these properties, the gated recurrent unit (GRU), long short-term memory (LSTM), and transformer-based LAG-LLAMA are trained and compared with an echo state network (ESN). The GRU, LSTM, and LAG-LLAMA models have not been successful in explaining the nonlinear oscillations of hydrogen bonding, obtaining insignificant values of R2 (0.05), whereas ESN has demonstrated outstanding predictive power (R2 = 0.99) with only 200 MD trajectory frames of initial information. In addition to this, despite the deep learning models achieving a high accuracy, via optimization, for mean-squared displacement (R2 = 0.95), the training sets are much larger (40-70%) than the ESN. In this way, the present work shows that reservoir computing models are more efficient and reachable in data efficiency and stability compared to the traditional recurrent and transformer models when applied to modeling the time-dependent transformation of complex molecular systems. This methodology will provide a generalizable methodology of fastening the computational design of soft materials bridging atomistic fidelity and AI-based temporal extrapolation.
Das et al. (Mon,) studied this question.