• Citer

Reduced-order modeling of supersonic fuel–air mixing in a multi-strut injection scramjet engine using machine learning techniques

par Ispir, Ali Can ;Zdybal, Kamila ;Saracoglu, Bayindir;Magin, Thierry ;Parente, Alessandro ;Coussement, Axel
Référence Acta astronautica, 202, page (564-584)
Publication Publié, 2022-11-10

Article révisé par les pairs

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Dual-mode ramjet/scramjet engines promise extended flight speed range and are the commonly preferred air-breathing propulsion system from within the family of hypersonic aircraft concepts. One of the main challenges that should be hurdled in their design is modeling the fuel–air mixing process to provide optimal fuel distribution and yield the best engine performance. Injecting fuel into high-speed air stream along multiple parallel struts can augment the fuel penetration and improve mixing efficiency. Mixing intensity is increased with turbulence by the shock-expansion waves in post-strut regions. However, this enhancement in mixing can bring about detrimental effects on the aerodynamic performance by increasing losses on total pressure. Designing the optimal working configuration requires testing the interaction between many design variables. This can be a tedious and computationally costly task. Machine learning models thus appear well-suited for multi-objective optimization of design variables that can be elusive to designers. In particular, non-linear regression models can be built over the available sparse simulation data to predict unseen mixing conditions. In the present work, we carry out a detailed investigation of the effect of multi-strut configuration parameters on three objective functions: mixing efficiency, mixing length, and the total pressure recovery (TPR) factor. These objective functions are linked with the most relevant physical phenomena in the supersonic fuel–air mixing flow field. We first generate a CFD database by solving compressible, non-reactive, Reynolds-averaged Navier–Stokes (RANS) filtered flow equations in a 2D scramjet engine domain with three varying design variables: struts location, strut wedge angle and strut V-settlement angle. We then apply various regression models – artificial neural network (ANN), Gaussian process regression (GPR), and kernel regression – to this database and formulate a surrogate model relevant for fuel injection that can be utilized in reduced-order modeling studies that estimate the hypersonic engine performance. We find that regression is generally more difficult in the vicinity of fuel struts (due to turbulence/shock effects) and easier further downstream from the struts, but ANN performs generally better than other regression models. Thus, our reduced-order tool incorporates a mixing efficiency model predicted by the ANN. It computes the thrust of a hypersonic engine with less than 10% error. We also present a detailed discussion of the physical insights gained from our CFD database; we link this discussion with the earlier findings from the machine learning tools. In our sensitivity study, we find the strut wedge angle to be the most influencing parameter on the mixing properties and aerodynamic losses.