Machine learning-aided reparameterization of a united atom model for chemically intricate polymer networks subjected to large tensile deformation

Efficient and accurate simulation of microscopic behavior and macroscopic properties of intricate polymer networks subjected to large tensile deformation is a challenging task for traditional coarse-grained (CG) and united atom (UA) models. In this study, we developed a machine learning functional calibration method to reparametrize a UA model for highly crosslinked and functionalized polymer networks subjected to substantial tensile deformation. The target material was a phosphorus (P) functionalized epoxy resin system, composed of Bisphenol A diglycidyl ether (DGEBA) and 4,4-Diaminodicyclohexylmethane (DDM) curing agent, which were functionalized by 10-(2,5-dihydroxyphenyl)-10-hydro-9-oxa-10-phosphaphenanthrene-10-oxide (ODOPB) functional groups. We constructed the calibration functional with nonbonded parameters as the calibrated parameters and densities (under different crosslinking degrees) and mechanical properties (within large tensile deformation range) as the targets. Two independent back propagation artificial neural networks (BP-ANNs) were trained and then combined, for density and mechanical property predictions, respectively, as the surrogate model to encapsulate the mapping relationship between the input calibration parameters and the output functional values. The multi-island genetic algorithm (MIGA) was employed to automatically determine the hyper-parameters of the BP-ANN, and also to seek out the optimal calibration parameters for the reparametrized UA (rUA) force filed The effectiveness and accuracy of the rUA model was validated, and the transferability of the model was examined to firstly predict tensile behavior of a similar material system with different weight ratio of P content, and then to predict materials densities and tensile mechanical properties under cryogenic temperatures (i.e., 90 K).