Multi-Fidelity-ML: A MATLAB repository from ShuaiGuo16

A Multi-fidelity Gaussian Process Approach to Identify Flame Model

1. Highlight

proposed a novel machine-learning-based strategy to effectively reduce the uncertainty in flame model identification, thus ensuring reliable combustor design and analysis.
This strategy relies on a multi-fidelity Gaussian Process model, which effectively aggregates low/high fidelity identification results and balances the trade-off between computational effort and accuracy.
We successfully tested the multi-fidelity strategy on the data acquired from a test rig.
Our results indicate that given the same computational budget, the proposed strategy yields globally more accurate, robust flame model identification.

This work was initially presented in CM4P ECCOMAS 2019 conference, and was later accepted in the journal:

Guo S., Silva C. F., Polifke W., Robust Identification of Flame Frequency Response via Multi-Fidelity Gaussian Process Approach. Journal of Sound and Vibration, 2021.

2. Motivation

Flame model constitutes a major source of uncertainty in combustion instability predictions. This uncertainty generally stems from the imperfect model identification from noisy time-series data. State-of-the-art identification methods are either accurate but terribly slow, or fast but contains significant uncertainties.

3. Methodology

We aim to fully exploit the respective strengths, while avoiding the weaknesses of the state-of-the-art methods by proposing a multi-fidelity machine learning approach to identify flame mdoels. This approach assimilates the global trend provided by the low-fidelity results and local estimates provided by the high-fidelity results, thus leading to a globally accurate and robust flame model identification even in the presence of strong noise.

4. Results

Given the same computational budget, our multi-fidelity strategy performs better than allocating resources solely to a single state-of-the-art method.
We extended our strategy to aggregate data from both simulations and experiments, demonstrating its flexibility in practical usages.

5. Folder structure

1. Presentation: the slides presented in CM4P ECCOMAS 2019 conference.

2. MatlabScripts: MATLAB source code and data to reproduce the results. The code and data are organized in individual folders corresponding to different sections in the paper.