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

Paper Number: 190377

190377 - Machine Learning Applications to Optimization of Composite Stiffened Panels 

The prevalence and accessibility of machine learning tools have created many opportunities in the structural optimization domain due to the size and complexity of problems that exist in this field. In particular, optimization of aerospace structures considers a very large design space due to the flexibility of modern manufacturing methods, especially composites. Despite constant improvements in computing power, Finite Element Model (FEM) complexity and size continue to grow at comparable rate, leading to challenges with runtime in structural optimization. This work considers applications of machine learning to structural optimization within HyperX, a commercial tool for aerospace structural optimization, in the pursuit of addressing these challenges. HyperX seeks the mass-optimum design for primary and secondary structures consisting of panel and beam components. In doing so, it must iterate through thousands or millions of different design candidates with varying thicknesses, stiffener geometry, and composite laminate stacking sequences. Each design candidate is evaluated against failure criteria for material strength and buckling across hundreds or thousands of different load cases. Navigating this vast constrained design space is possible but can require experience to arrive at a usable solution.  In the present study, two categories of machine learning are being considered for application in HyperX. The first is to create models for the structural failure criteria called by HyperX to evaluate each design candidate. This results in a surrogate of the optimization constraints that can return the constraint function (margin of safety) much faster than the original analysis method. The second category of application is to construct a surrogate of the optimization itself. In this scenario, the model predicts the optimum design (dimensions and or composite stacking sequence) for the stiffened panel or beam based on the loads in the structure and the material properties. Both categories are investigated simultaneously due to the advantages and limitations they each present. The first category provides more flexibility in the optimization, because it is common for different failure analyses to be used in different problems. The second category has reduced flexibility because the failure analyses are “baked-in” to the model. However, this approach is faster than the first because it predicts the optimum design in one shot without any iteration. This work explores the development of models in these two categories and examines the feasibility of applying them to structural optimization in a tool such as HyperX.

Presenting Author: August Noevere Collier Aerospace

Presenting Author Biography: August is the Director of Research at Collier Aerospace, where he divides his time between research programs, development of HyperX software, and supporting customer projects as an analyst. His primary interests are in design and optimization of composite structures and digital thread integration.