The research focuses on developing methods and models for space structures manufactured using novel production techniques. The initial activities concentrated on modeling lattice structures produced through additive manufacturing. These structures are formed by the repetition of a representative volume element, or unit cell, which consists of multiple interconnected struts.

Because of the very small dimensions of these components, the manufacturing process introduces geometric and material imperfections. These include porosity, irregular surfaces, and geometric inaccuracies, all of which significantly influence the structural response. Accurately accounting for these defects is therefore essential to obtain reliable predictions of mechanical behavior.

To address this challenge, a numerical framework has been developed to generate detailed three-dimensional finite element models of imperfect lattice structures. The imperfections are introduced through an automated random procedure, enabling the generation of realistic geometries. Although these models provide high fidelity, they are computationally expensive and therefore limited to analyses at the cell scale.

To overcome this limitation, simplified models based on one-dimensional beam or truss elements have been developed. In these models, the mechanical response of an imperfect strut, obtained from the detailed three-dimensional simulations, is used as the constitutive law of the corresponding one-dimensional element. In this way, the effects of manufacturing imperfections are embedded in the structural response while maintaining a low number of degrees of freedom. This approach enables the analysis of full-scale lattice structures with significantly reduced computational cost.

A neural network has been developed and trained using data from the high-fidelity finite element simulations in order to automate the assignment of constitutive laws to the one-dimensional elements. The network takes as input the geometric characteristics of a strut together with parameters describing the imperfections, and outputs the corresponding constitutive law and section properties.

This framework enables fast and accurate assessment of lattice structures, accounting for manufacturing-induced defects. It can be used to analyze large-scale structures and to evaluate multiple configurations for optimization purposes.

Current research activities focus on developing an artificial intelligence–based optimization procedure for lattice structures. Two-dimensional lattices are generated and modeled using the reduced-order framework, and the resulting data are used to train a graph neural network encoder that acts as a surrogate model. Optimization is performed in the latent space of the encoder, and a decoder reconstructs the optimized lattice configuration. Once validated, the methodology will be extended to three-dimensional lattices.

Click here to view my poster.