Physics-Informed Neural Operators for Generative Design of Tuned Vibration Absorbers in Urban Air Mobility

The primary goal of this project is to develop a data-driven design assistant for generative design of vibration absorbers with application to eVTOLs. Recent results demonstrate promising potential for integrating operator learning with conventional numerical methods in design optimization tasks. Herein, a neural operator is trained to learn the inverse mapping between a desired target spectrum and corresponding design parameters. This approach eliminates the need for an iterative optimization process, which is traditionally required to adjust parameters until the simulated frequency response matches the target spectrum. Instead, the trained neural operator can directly predict the necessary design parameters for a given target spectrum. The neural operator acts as a surrogate model. After training, this operator can generalize new designs that were not seen during training, making the method a generative design algorithm. It provides engineers with a strong initial estimate for parameters, which can then be refined through high-fidelity simulations to further explore optimal design regions.

The process begins with the forward simulation model, which generates training data pairs by varying design parameters and obtaining the corresponding transmission loss spectra. Once trained, the neural operator proposes novel tuned vibration absorber designs tailored to a specified transmission loss spectrum. This innovative approach offers several advantages: Operator learning facilitates real-time generative designs, providing engineers with fast, efficient, and reliable assistance throughout the design process. Neural operators facilitate data fusion during training, allowing both simulation and experimental data to be incorporated into the model, representing a novel approach to model fusion. This approach leads to more realistic models that help bridge the gap between simulation and experiment, eliminating the need for tedious parameter tuning until simulation matches the experiment. Additionally, this data-driven design assistant facilitates interactive design, allowing engineers to directly observe the effects of design changes through rapid evaluation performed by the neural operator. This enables them to make quick adjustments as needed.

• Development of a data-driven design assistant for vibration absorbers: Unlike traditional design processes, this project aims to develop a data-driven design assistant for tuned vibration absorbers, enabling a direct, objective, and interactive design process.
• Implementation of neural operators as a data-driven surrogate model: Neural operators are implemented and trained to directly map between a target variable and design parameters.
• Integration of physics information into the neural operator model: Known physical equations of mechanics and elastodynamics will be incorporated into the loss function of neural operators to enhance prediction accuracy. This integration of physics knowledge improves the model’s performance by constraining the learning process with known laws of physics.
• Generative design and a new design paradigm: Once the physics-informed neural operators are trained, the model predicts designs beyond the training dataset, enabling the generative design of new vibration absorbers. This results in a novel design paradigm, where the design assistant supports engineers by providing fast, interactive, and robust solutions.

• Model Order Reduction Techniques
• Multi-Objective Optimization Strategies

• Numerical Methods such as Finite Element Modeling and Boundary Element Method
• Vibro-Acoustic Analysis and Solutions
• In future: Physic-Informed Operator Learning Solutions, Uncertainty Quantification