A machine-learning-based peridynamic surrogate model for characterizing deformation and failure of materials and structures

It is necessary to determine the input features and output results when constructing a surrogate model within the data-driven neural network. Since the law of features would be restrained when the surrogate mechanical model is employed, it is still a challenge to build a set of natural features to accurately describe the failure process of materials and structures within the traditional continuum mechanics framework. To address this challenge, a robust approach for constructing a surrogate model within the peridynamic-deep learning framework is proposed in this study, which is capable of representing material deformation and failure explicitly. The presented surrogate model integrates both reference and current configuration data to refine input features, enhancing model training. We incorporate a batch-normalization layer before the activation function to mitigate common issues such as slow convergence, low prediction accuracy, and overfitting due to the large numerical differences in the damage dataset. Additionally, numerical analyses on several typical examples are performed to validate the effectiveness and generality of the present model and methodology. The results demonstrate high accuracy in the training set as well as the testing set, confirming the model’s excellent generalization ability and significant potential for material failure analysis. According to this work, more peridynamic expressions can be further derived in the machine-learning-based peridynamic surrogate model by considering the reinforcement learning and symbol space, to potentially broaden its applicability to a wider range of mechanical issues.