REML - Reliable Machine Learning for Thermoacoustic Modeling

The move toward hydrogen-based combustion in gas turbines and aircraft engines is a key step in reducing greenhouse gas emissions. While hydrogen offers a promising path toward cleaner propulsion, developing these systems is far from straightforward. One of the biggest challenges is thermoacoustic instability—a harmful feedback loop between flame dynamics and acoustic waves that can create high pressure oscillations, potentially damaging or destroying an engine.

Today, engineers often rely on computational fluid dynamics (CFD) to analyze these phenomena. Although CFD provides detailed insights, it is computationally expensive and slow, limiting its use during fast design and testing cycles.

This is where machine learning (ML) comes in. ML has become a powerful tool across engineering, and it can help accelerate predictions of thermoacoustic behavior using existing data. However, thermoacoustics data are usually sparse, noisy, and difficult to obtain, which raises concerns about the reliability of purely data-driven predictions—especially in safety-critical systems like airplane engines.

Our research focuses on making ML trustworthy and useful for real thermoacoustic problems. A major theme is uncertainty quantification (UQ): understanding not only what a model predicts, but also how confident it is.

We approach this through four main research directions:

1. Bayesian Neural Networks (BNNs) for flame dynamics
We use BNNs to model nonlinear flame responses. Unlike standard neural networks, BNNs provide natural measures of uncertainty stemming from limited or noisy data.

2. Data value and experimental design
By analyzing BNN uncertainties, we identify which data are most informative for learning flame dynamics—helping guide smarter, more efficient experiments/simulations.

3. Physically inspired colored combustion-noise models
We develop models for correlated (colored) combustion noise to help separate signal from noise. This improves the robustness and interpretability of BNN-based predictions.

4. Evidence-based model selection
Using Bayesian evidence, we automatically determine which BNN architecture is most plausible for a given task—avoiding trial-and-error model tuning.