A comparison of classical phenomenological, hybrid neural network, and high performance ROM FE2 models for open-cell foams: Efficiency, accuracy, and flexibility

Computational homogenization has become an attractive path to determine the macroscopic inelastic behavior of a structure from the knowledge of the behavior of its (usually simpler) constituents, using a Representative Volume Element (RVE) at the microscale. Since the numerical solution of both scales by means of the FE² method is prohibitively expensive, much research effort has been spent on the development of surrogate models, particularly by making use of data-driven machine learning methods. Due to the vast spectrum of proposed surrogate modeling approaches, the decision on the method of choice for potential applicants is difficult. While computational efficiency is usually used as the central performance measure, a careful distinction between on- and offline efforts is important. Moreover, the simulation context should also influence the selection. If, for instance, a thousand simulations are to be carried out for a single microstructure, the choice of approach may be different than if single, or a few, simulations are to be carried out for a thousand different microstructures. In the latter case, the flexibility and adaption effort of the method to new microstructures plays a key role. The present study aims to derive guidelines in this direction, by comparing different such approaches and carefully analyzing their performance in simulating inelastic macroscale behaviors, influenced by the morphology of 3D foam-like microstructures. Four formulations are considered: the Deshpande–Fleck phenomenological macroscale model, a general analytical surrogate model, a hybrid neural network surrogate model, and a hyper-reduced FE² approach. The methods are compared with respect to computational costs and predictive capabilities for two 3D benchmark scenarios.