Session: 12-08-01: Data-Enabled Predictive Modeling, Scientific Machine Learning, and Uncertainty Quantification in Computational Mechanics

Paper Number: 149680

149680 - End-to-End Reduced-Order Modeling in Simulink 

This talk offers a comprehensive overview of how artificial intelligence can be leveraged to accelerate the simulation of complex, high-fidelity models, such as those derived from Finite Element Analysis (FEA), Computational Fluid Dynamics (CFD), and Computer-Aided Engineering (CAE). The presentation focuses on the Model-Based Design approach, emphasizing the significant role AI can play in component modeling and algorithm development. High-fidelity models, known for their computational intensity, can be effectively simplified using AI-driven ROMs, thus making simulation and analysis more efficient.

Reduced-order modeling simplifies full-order high-fidelity models by reducing their computational complexity while preserving their dominant behavior. This simplification enables the integration of third-party FEA/FEM/CFD models into system-level simulations within Simulink. By replacing complex component-level models with ROMs, engineers can perform real-time simulations, including hardware-in-the-loop (HIL) testing. This is particularly beneficial for control algorithm testing on embedded hardware, where the reduced computational load makes real-time execution feasible.

ROMs can also serve as virtual sensors, offering an alternative to physical sensors when direct measurement is impractical or impossible. This capability is invaluable for estimating or predicting signals of interest in various engineering applications. Additionally, ROMs facilitate control design by simplifying the complexity of the plant models, making it easier to design and validate controllers. For advanced control algorithms like nonlinear model predictive control, reduced order modeling can be used to create an internal prediction model when it’s hard to obtain a first-principles model.

The presentation includes an example of creating a reduced-order model of a jet engine turbine blade. This example utilizes Long Short-Term Memory (LSTM) and neural state-space (NSS) models to capture the turbine blade's maximum displacement under varying ambient temperature, cooling temperature, and pressure conditions. The workflow begins by using the Reduced Order Modeler app to specify the inputs and outputs for the ROM, design experiments, and collect data from the high-fidelity model. The collected data is used to train an LSTM model, which captures the system dynamics from the input conditions to the maximum displacement. We train various models and identify the best-performing ones by sweeping over different hyperparameters.

Similarly, an NSS model is created to capture the system dynamics. We train the NSS model with different hyperparameters and evaluate the model performance. The best-performing NSS model is also exported to the MATLAB workspace. Both the LSTM and NSS models are then integrated into a Simulink model, where their performance is validated against the high-fidelity model using independent validation data. The trained models show good generalization and can be further refined by adding more experiments or tuning hyperparameters. Ultimately, the identified ROMs can be used for HIL testing and control design in Simulink or exported for use outside Simulink through Functional Mock-up Units (FMUs).

Presenting Author: Mehdi Vahab MathWorks

Presenting Author Biography: Mehdi Vahab is the Academic Manager for Mechanical and Aerospace Engineering at MathWorks. His academic background is in physical modeling, specifically for fluid and thermal systems, and applications of computational modeling in different engineering problems. Before joining MathWorks, he developed numerical methods for the computational modeling of multiphase/multimaterial systems and phase-change dynamics throughout his research career. He applied those methods to study scientific and engineering problems, such as using active vortex generators to manipulate heat transfer in pool boiling, thermal management of hypersonic vehicles with heat pipes, and understanding how the melting process enhances heat transfer when snowing in open waters. At MathWorks, he helps researchers, faculty, and students in Mechanical and Aerospace Engineering with their research/teaching problems by collaborating and consulting to find better and more accessible solutions.

Authors: