Graph neural network-driven surrogate modeling for accelerated parameter identification in semicrystalline polymers

Accurately identifying material parameters in phenomenological models is crucial for capturing the complex behavior of semicrystalline polymers. However, the iterative nature of optimization and probabilistic algorithms within Finite Element Analysis (FEA) often makes this process computationally intensive and time-consuming, especially when high-fidelity constitutive models are involved. To address these challenges, we developed a surrogate model based on graph neural networks (GNNs) to accelerate the parameter identification process for semicrystalline polymers. GNN surrogate models excel in applications where repetitive simulations are required using the same underlying model, such as parameter identification, design optimization, and sensitivity analysis, while inherently enabling full-field predictions and flexibility in handling input nodes, ranging from sparse experimental measurements to dense full-field datasets. The GNN was trained on data generated from a constitutive viscoelastic, viscohyperelastic, and viscoplastic model implemented in FEA. The surrogate model achieved high accuracy across a wide range of parameters, although larger errors were noted near the boundaries of the training set and during extrapolation. Despite these limitations, by approximating the input–output relationships of computationally intensive FEA simulations, GNNs significantly enhance speed and efficiency. The GNN outperformed FEA computations in terms of speed, delivering results in under 9 s, compared to the 4 plus hours typically required by traditional methods. This significant improvement in computational efficiency is critical for industrial applications, such as digital manufacturing, where reducing model optimization time can accelerate the development of new materials and technologies.