ERI: Physically Constrained Generative Artificial Intelligence for Feasible Space Transformation in Design Optimization

This Engineering Research Initiation (ERI) project support research that aims to enable rapid engineering design optimization involving complex design requirements. Engineering design is an approach for optimizing objective(s) (e.g., minimizing the drag of an airplane) while fulfilling design requirements (e.g., maintaining a lower limit of the lift generated by the airplane). Engineering design has achieved success in broad areas, including automobiles, aircraft, spacecraft, energy systems, and manufacturing equipment. Conventional design optimization, however, faces the following two challenges in practical applications. 1) Objective values and design requirements typically need to be computed hundreds or even thousands of times, which significantly slows down the optimization process. 2) Practical optimization typically involves complex design requirements, which make it difficult to find real optimal solutions. To overcome these challenges, this project will investigate developing an approach that ensures all design requirements are automatically fulfilled during the optimization process. If successful, this approach can reduce optimization complexity and increase optimization efficiency. This project seeks to develop a generative artificial intelligence approach for design requirement fulfillment, along with surrogate models to enable near real-time computation – ultimately supporting rapid engineering design. Generative artificial intelligence can learn from existing data (e.g., takeoff trajectories of electric drones) and generate similar data patterns (e.g., smooth takeoff trajectories without sudden drops). Surrogate models, also known as predictive models, are fast approximations of time-consuming mathematical models. In this project, generative artificial intelligence models will be guided by surrogate model predictions to assess whether design requirements are fulfilled by the generated designs. This feedback loop allows the generative artificial intelligence to generate designs that inherently fulfill all design requirements during optimization. Performance of the developed approach will be evaluated using an optimal takeoff trajectory design of an electric drone verified using NASA’s Dymos framework, as well as on two aerodynamic design benchmark problems developed by the AIAA Aerodynamic Design Optimization Discussion Group. In summary, this research project seeks to enable rapid engineering design under complex design requirements while maintaining high accuracy on the resulting optimal solutions. Moreover, this project aims to discover new knowledge about how design requirements can be effectively fulfilled using generative artificial intelligence and surrogate models. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.