Session: 02-03-02: Optimization

Paper Number: 144578

144578 - A Data-Driven Design Pipeline for the Structural Optimization of Quadcopters 

The application of unmanned aerial vehicles (UAVs) in extreme situations, such as disaster rescue or military reconnaissance, has become increasingly prevalent. The increasingly hazardous nature of these tasks warrants consideration of optimization techniques to improve structural integrity. To optimize quadcopter parameters, designers currently make incremental, intuitive design advancements. The iterative design process is often tedious, time-consuming, and expensive, detracting energy from creative insight. Generative design and data-driven engineering are appealing solutions that automate optimization based on specified user requirements. Applying such techniques to quadcopter optimization is essential to tackle a new forefront of challenges. This research paper uses data-driven design, performance prediction models, and AI-driven design recommendation tools to structurally optimize quadcopters (a subset of UAVs). The presented methodology offers a design assistant for human designers, satisfying rigorous design constraints through a streamlined and replicable process.

The methodology involves first creating a fundamental quadcopter model with nine parameters (width of drone, width of arm, arm angle relative to vertical, arm taper angle, width of body, length of drone, arm thickness, and arm shell thickness) representing ubiquitous features across quadcopters. Thousands of models are generated and eliminated via geometric and model infeasibility checks. 3D models of designs are built using SolidWorks, and finite element analysis (FEA) is used to analyze the structural performance of designs to understand behavior under various physical conditions. The output of FEA is four structural performance values (mass, minimum factor of safety, displacement, and von Mises stress). An Automated Machine Learning (AutoML) regression model is trained on 5000 models to predict the above created. Model accuracy is validated on each structural performance value utilizing metrics such as r-squared and mean-squared error, prediction, and residual plots. To analyze the sensitivity of performance parameters with respect to design parameters, a SHapley Additive exPlanations  (SHAP) is performed. To simultaneously optimize multiple queries, Multi-Objective Counterfactuals for Design (MCD), which utilize inverse counterfactual thinking, recommend specific modifications. 

Initial findings suggest that the results from SHAP and MCD provide several key insights that are immediately applicable to the design process. The relationship between design parameters and structural performance values is quantified through SHAP. This includes determining which parameters have an outsized or undersized effect on structural performance values and the ability to optimize a performance value without significantly affecting another. Furthermore, several case studies with stringent design parameters are conducted to demonstrate the efficacy of MCD. Both single-objective and bi-objective queries are inputted. Further FEA testing indicated that all outputted design recommendations fulfilled the desired queries. 

This work establishes a novel data-driven approach to parametric analysis and optimization of quadcopter designs. It introduces a parameterized dataset of thousands of different models, quantifies relationships between parameters and performance values, and presents immediate design recommendations. It is a valuable companion to human designers because it rapidly discerns patterns, leaving more time for creative insight. Additionally, compared to other techniques, such as topology optimization, this method gives designers more control over individual steps, from parameterization to design recommendations. Due to the replicable nature of the methodology, it can be applied outside the scope of structural optimization (i.e., aerodynamic or aesthetic optimization). 

 

Presenting Author: Rishab Subramanya Energy Institute High School

Presenting Author Biography: Rishab Subramanya is a junior at Energy Institute High School in Houston, Texas. His interests include data-driven design, optimization technology, and machine learning.

Authors:

Rishab Subramanya Energy Institute High School