Multifidelity computing for coupling full and reduced order models

Area of Science:

• Computational Science
• Machine Learning Applications
• Multiphysics Simulations

Background:

• Hybrid physics-machine learning models are crucial for simulating complex transport processes.
• Multiphysics systems often involve multiple spatiotemporal scales and multifidelity formulations.
• Integrating diverse computational entities presents a significant challenge in scientific and engineering applications.

Purpose of the Study:

• To develop a robust hybrid analysis and modeling approach for integrated mixed-fidelity simulations.
• To create building blocks for predictive digital twin technologies by combining full-order and reduced-order models.
• To introduce an interface learning method for effective coupling of high and low-fidelity models.

Main Methods:

• A physics-based full-order model (FOM) was combined with a data-driven reduced-order model (ROM).
• A long short-term memory (LSTM) network was employed at the interface to bridge FOM and ROM.
• The approach was tested on nonlinear advection-diffusion flow problems in a bifidelity setup.

Main Results:

• The proposed interface learning effectively corrects or prolongs information between mixed-fidelity models.
• The hybrid approach successfully addresses ROM-FOM coupling challenges in transport process simulations.
• The method demonstrates a new way to handle interfacial errors in complex simulations.

Conclusions:

• Hybrid physics-machine learning models offer a powerful framework for advanced simulations.
• LSTM networks provide an effective mechanism for interfacing different fidelity models.
• This integrated approach advances the development of predictive digital twin technologies for transport processes.