Quesada-González C., Dehghani H., Belouettar S., Penta R., Ramírez-Torres A., Merodio J.
Mechanics of Composite Materials, 2026
A data-driven approach is developed, as an upscaling means for composite media to overcome the need to fully resolve the complex cell problems encountered in asymptotic homogenization. Artificial neural networks (ANNs) were applied as an instantaneous (black-box) transformation means from the input space to the output space. The former comprises the volume fraction and material properties of each microscale constituent, while the latter provides the effective properties of the homogenised model. The ANN training utilizes a training dataset that provides a sufficiently dense discretization of a continuous input-output function through sampling approaches. The deterministic low-discrepancy sequences were used to provide optimal sampling of multidimensional input space. The corresponding outputs in the dataset were generated using a semi-analytical approach. Our results indicate that the ANN achieves high accuracy. Additionally, we analyze how the size and density of the training dataset impact the accuracy of the upscaling, providing insights into the computational efficiency and applicability of data-driven methods in multiscale analyses.
