Multi-Fidelity Surrogate Modeling for Unmanned Combat Aerial Vehicles

My research focuses on how combining different levels of model fidelity can greatly enhance the efficiency and accuracy of solving complex engineering problems. One of my key projects involved the use of multi-fidelity surrogate models to optimize the design of unmanned combat aerial vehicles (UCAVs). In this approach, I blended high-fidelity simulations—such as computational fluid dynamics (CFD) results from dense grids, which offer precise but computationally expensive solutions—with low-fidelity models like CFD results from coarser grids that provide faster but less detailed outputs. By employing hierarchical Kriging and adaptive sampling techniques, the multi-fidelity framework was able to capture a wide range of aerodynamic characteristics across various flight conditions with much lower computational costs. This strategy allowed for a broader exploration of design spaces, offering more opportunities to fine-tune the design and gain valuable insights into the complex relationships between different design variables and their impact on performance metrics.

A significant outcome of this work was the use of the trained multi-fidelity model to augment the datasets, providing a robust platform for extensive exploration of design spaces. With these enhanced datasets, I employed various data mining techniques, including decision trees and rough set theory, to identify critical design parameters that govern the UCAVs' aerodynamic performance. This comprehensive approach tackled multiple objectives, addressing both low-speed stability and high-speed aerodynamic efficiency. As a result, it created a more balanced and effective model capable of navigating the diverse and often conflicting demands inherent in UCAV design optimization, demonstrating the power and flexibility of multi-fidelity modeling in real-world engineering applications.

Click for more details

Multi-Fidelity Modeling for Physics-Informed Neural Networks

Expanding on this, my next work integrated multi-fidelity modeling with physics-informed neural networks (PINNs) to enhance digital twins. This integration allowed for real-time predictions while accounting for uncertainty, combining both data-driven and physics-based models. By leveraging multi-fidelity datasets, the digital twin became capable of accurately simulating a variety of physical conditions, making it far more adaptable than conventional methods.

The use of multi-fidelity models enabled the digital twins to provide updated virtual representations efficiently. This approach significantly increased prediction accuracy while maintaining a similar computational cost, allowing the digital twins to handle a wide range of complex scenarios. The added ability to incorporate data of varying fidelity levels ensures that the digital twin remains robust, even when dealing with sparse or heterogeneous datasets common in engineering applications.

Click for more details