Session: 01-05-02: AI-Driven Modeling and Simulation for Aerospace Structures

Paper Number: 189001

189001 - Multiscale Analysis of Sandwich Structures With Meta-Cores Using a Transfer Learning Approach 

Sandwich structures are extensively used in aeronautical, naval, and automotive applications due to their high stiffness-to-weight ratios, structural efficiency, and multifunctional capabilities. With recent advances in additive manufacturing, sandwich cores with architected microstructures can now be fabricated with unprecedented geometric complexity, enabling tailored mechanical responses. However, this expanded design flexibility significantly increases the computational burden of analysis. Finite element based direct numerical simulation (DNS), while accurate, becomes prohibitively expensive for large sandwich structures with complex core geometries, particularly when detailed microstructural resolution is required.

To address this challenge, we employ a multiscale modeling framework based on the Mechanics of Structure Genome (MSG) theory for efficient structural analysis of sandwich plates with metamaterial cores. In MSG, the Structure Genome (SG) represents the smallest mathematical building block that completely characterizes the material’s microstructural topology and constituent distribution. MSG establishes a mathematically consistent linkage between the SG and macroscopic structural behavior, enabling direct computation of effective stiffness and structural response without empirical assumptions.

With the MSG framework, two multiscale plate analysis strategies are systematically investigated. The first strategy employs a two-step approach. In the first step, the effective properties of the core metamaterials are obtained based on MSG homogenization analysis. The second step uses one-dimensional (1D) SG to describe the facesheets and the core materials with effective properties, efficiently computing the entire effective plate stiffness for structural analysis. This approach significantly reduces the computational costs compared to DNS with good accuracy. However, this approach requires the core materials containing multiple unit cells in the plate thickness direction to meet the scale separation assumption (SSA). In practice, volume constraints and manufacturing resolution often limit sandwich structures to only one or two unit cells across the thickness, which violates the SSA and leads to significant loss of accuracy. To address this issue, the second strategy uses a one-step approach in which the entire core and face sheets are treated collectively as a one SG and analyzed directly to compute the entire plate stiffness. Although more computationally intensive than the two-step method, the one-step approach remains significantly less expensive than full DNS while maintaining high accuracy without SSA.

The proposed strategies are demonstrated using sandwich structures with re-entrant honeycomb (2D SG) and body-centered cubic lattice (3D SG) metamaterial cores. Comparative studies against DNS show that the one-step approach achieves near DNS accuracy across all configurations, while the two-step approach offers a highly efficient alternative when sufficient periodicity exists in the thickness direction. These results establish clear accuracy efficiency tradeoffs between the two MSG strategies and provide practical guidance for selecting appropriate multiscale modeling approaches based on core architecture and thickness.

To further enable rapid exploration of the expanded design space, we develop a multi-fidelity transfer learning framework for predicting effective stiffness properties of sandwich structures with re-entrant honeycomb cores. Low-fidelity data generated using the efficient two-step MSG approach are used to capture global trends across a broad parameter space. A limited set of high-fidelity data based on one-step MSG method is then employed to fine-tune the model. This multi-

fidelity strategy significantly reduces the required high-fidelity data and overall training cost while maintaining strong predictive accuracy.

Overall, this work presents two efficient MSG-based multiscale strategies for sandwich structures with metamaterial cores and introduces a multi-fidelity transfer learning framework that substantially accelerates accurate design space exploration.

Presenting Author: Arnob Nandy Mazumder University of Texas at Arlington

Presenting Author Biography: Arnob Nandy Mazumder is a PhD student in the Mechanical and Aerospace Engineering department at the University of Texas at Arlington. His research focuses on machine learning assisted multiscale modeling of sandwich structures.