ResNet-assisted coarse-grained modeling of CH4-CO2 hydrates: Linking structural descriptors to interaction potentials.

Abstract

Gas hydrates are promising for energy storage, gas separation, and carbon sequestration. However, atomistic simulations of hydrate systems are computationally expensive at large scales. In this work, a machine-learning-assisted coarse-grained (CG) force field is developed for CH4-CO2 hydrate systems based on the Stillinger-Weber (SW) potential. A deep residual convolutional neural network (ResNet) is employed to establish a nonlinear mapping between structural descriptors and cross-species SW interaction parameters. The trained model achieves high accuracy (R2 > 0.99) and reproduces key structural and thermodynamic properties of the reference atomistic systems, including radial distribution functions, diffusion coefficients, interfacial tension, and gas solubility. With the enhanced computational efficiency of the CG model, large-scale simulations were conducted to explore hydrate dissociation and defect-controlled mechanical behavior. Results reveal distinct bubble nucleation and growth characteristics among CH4 hydrate, CO2 hydrate, and mixed hydrate systems, arising from differences in gas-water interactions. Structural defects, particularly in the water framework, weaken the hydrate lattice most significantly. The ResNet-assisted parameterization method provides an efficient strategy for constructing CG force fields for multicomponent hydrate systems, enabling simulations at scales beyond conventional atomistic methods.