Date of Award
8-2026
Document Type
Thesis
Degree Name
Master of Science (MS)
Department
Mechanical Engineering
Committee Chair/Advisor
John R. Wagner
Committee Member
Cameron Turner
Committee Member
Laura Redmond
Abstract
Engineering design efforts benefit from a broad range of system knowledge to form a basis for informed decision-making. Technical information gained from experimentation, virtual tools, and analytical methods can be applied early in the design phase to reduce the need for time-consuming engineering changes later in development. Therefore, it is in the best interest of designers and decision-makers to leverage the proper engineering tools to complete the product discovery task efficiently. In complex, large systems with multiple subsystems, design choices must balance a multitude of objectives. Trial-and-error design methods that explore decision alternatives and the potential of different designs are not time-efficient and rely on subject-matter experts, predefined rules, and/or extensive experience. The tradeoffs among alternative designs, mapped in the tradespace, are highly dimensional as metrics continually grow with system complexity, making them difficult to navigate and visualize purposefully. Recent advancements in digital technologies, such as virtual agents, digital twins, and machine learning, enable the creation of a process that efficiently navigates the tradespace for embodiment design, automating portions of this activity. By applying powerful artifacts, designers can gain key insights, streamlining product discovery for multi-objective design.
The thesis research objective is to explore the viability of a digital agent-based framework that utilizes emerging next-generation intelligent technologies for subsystem design. The agent-based strategy integrates machine learning prediction algorithms, called predictors, with a digital twin to enable direct mapping from performance-space targets to feasible design configurations. Predictors are trained for each design variable using a dataset of bulk simulations generated by the digital twin. In contrast to traditional iterative approaches to design and optimization, which require convergence through repeated evaluations, the agent identifies regions of the design space that satisfy specified performance objectives without iterative search. This is enabled by machine learning predictors trained on the entire design and performance spaces. The proposed methodology is demonstrated through a case study involving the design of an off-road tracked vehicle powertrain subsystem. Three Deep Orange 14 off-road tracked vehicle powertrain variants, defined by distinct gear ratios, are explored using a consistent set of design variables and performance metrics. Additionally, a comparison between the agent strategy and a traditional genetic algorithm is completed to determine their relative abilities.
Numerical results show that the digital agent generated configurations that closely match target performance while adapting to each variant's capabilities. The best design produced by the agent for the three powertrain variants had RMS percent deviations of 4.7%, 1.8%, and 2.3%, respectively. The best solution offered by the digital agent was from variant B, with the following design variables: battery capacity = 37.0 Ah, electric motor torque = 97.9 Nm, electric motor power = 37,603 W, aerodynamic drag coefficient = 1.55, (higher than expected) and viscous rolling resistance coefficient = 9.44*10-3 s/m. A comparative study with a genetic algorithm highlights differences in solution strategy, computational efficiency, and design outcomes. The genetic algorithm used MATLAB’s NSGA-II variant and was applied to powertrain variant B. The best design discovered during the algorithm’s search had an RMS percent deviation of 2.0% with battery capacity = 46.2 Ah, electric motor torque = 98.9 Nm, electric motor power = 46.532 W, aerodynamic drag coefficient = 1.09, and viscous rolling resistance coefficient = 1.32*10-2 s/m. The digital agent generated training data and designs in 18 hours, while the genetic algorithm took 64 hours to complete 100 generations. Overall, the agent navigated the tradespace more efficiently; however, the genetic algorithm created a more diverse Pareto frontier. The agent-based strategy is a data-driven engineering tool that assists human designers in embodiment design by exploring continuous, multi-objective, and high-dimensional design and performance spaces through tradespace exploration. These results demonstrate the potential of agent-based, data-driven strategies to aid human decision-makers in embodiment design and tradespace exploration of complex systems.
Recommended Citation
Bradley, Andrew J., "Intelligent Digital Agents in Mechanical Design With Application to Transportation Systems" (2026). All Theses. 4835.
https://open.clemson.edu/all_theses/4835