From abstraction to design: Interpretable tree-based machine learning for stable thermoacoustic system layout

Thermoacoustic instabilities are a long-standing problem. Both academia and industry researched intensively this problem, but even today, no guidelines (or design rules) exist to layout a stable combustor. It is often unclear how the system parameters, individually, and in combination, contribute to the thermoacoustic stability. Knowing these relationships would provide useful insight into the problem domain and accelerate the design process. This paper introduces a novel workflow to identify the key functional relationships between the design parameters and the stability. The workflow consists of (i) combining acoustic network models and genetic optimization to generate a rich data set that links system parameters (e.g., reflection coefficients, geometrical extensions, and flame dynamics) to the system stability (the growth rate of the eigenmodes); (ii) using this data set to train tree-based machine learning models for stability prediction given a set of design parameter; (iii) applying interpretable machine learning, namely SHAP values, to identify the key information of which and to what extent the design parameters affect the combustor stability. Finally, we demonstrate, how this workflow successfully helps to redesign an unstable reference system configuration to a stable one.