CAREER: Single-Fidelity vs. Multi-Fidelity Computer Experiments: Unveiling the Effectiveness of Multi-Fidelity Emulation

Michigan State University, East Lansing MI

Investigators

Abstract

Computer models have become indispensable tools across diverse fields, enabling the simulation of complex phenomena and facilitating decision-making without costly real-world experiments. Traditionally, computer models are simulated using single, high-accuracy simulations, employing a high level of detail and resolution throughout. Recent advancements, however, have shifted attention towards multi-fidelity simulations, balancing computational cost and accuracy by leveraging various levels of detail and resolution in the simulation. A key question arises: is it more effective to use single-fidelity or multi-fidelity simulations? This is a question practitioners often confront when conducting computer simulations. The research aims to address this fundamental question directly, providing valuable insights for practical decision-making. By leveraging insights gained from computational cost comparisons, the research will enhance the ability to predict complex scientific phenomena accurately and has the potential to revolutionize fields such as engineering, medical science, and biology. The project contributes to outreach and involvement in the REU exchange program, provide opportunities to engage undergraduate students, nurturing their interest in research and encouraging them to pursue careers in STEM. Research findings will be disseminated through publications and conferences. The code developed will be shared to foster collaboration and encourage others to build upon these innovative methodologies. This research addresses the fundamental question of whether to conduct single-fidelity or multi-fidelity computer experiments by investigating the effectiveness of multi-fidelity simulations. It begins by examining the computational cost comparison between the two approaches, finding that multi-fidelity simulations, under certain conditions, can theoretically require more computational resources while achieving the same predictive ability. To mitigate the negative effects of low-fidelity simulations, a novel and flexible statistical emulator, called the Recursive Nonadditive (RNA) emulator, is proposed to leverage multi-fidelity simulations, and a sequential design scheme based on this emulator is developed, which maximizes the effectiveness by selecting inputs and fidelity levels based on a criterion that balances uncertainty reduction and computational cost. Furthermore, two novel multi-fidelity emulators, called "secure emulators," are developed, which theoretically guarantee superior predictive performance compared to single-fidelity emulators, regardless of design choices. 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.