Multi-fidelity Strategies to Support Pintle Injector Design
Please login to view abstract download link
The pintle injector represents a promising solution for deep throttling in liquid rocket engines. The large range of working conditions in which it can operates introduces several design challenges. In particular, the numerical prediction of its performance in a wide range of conditions is computationally demanding. For these reasons, it is important to develop Reduced Order Models (ROMs) to efficiently predict its performance. In this work, the multiphysics nature of the problem is investigated by developing several numerical models for some basic flows which represent the different phenomena which control pintle performance. Specifically, some canonical test cases characterized by large separations and sprays are selected from the literature and investigated by means of Reynolds Averaged Navier-Stokes (RANS) equations for different working conditions. The obtained database is used to train different classes of ROMs. First of all, surrogate models based on multi-fidelity strategies [1] are used to evaluate global quantities like reattachment position or spray angle. The multi-fidelity approach allows to combine data from different levels of accuracy: in this work, a large database of cheap RANS solutions obtained on a coarse grid is used as low fidelity data while high-fidelity data are available from a smaller database of expensive RANS simulations performed on a fine grid. Different multi-fidelity strategies are compared and evaluated changing the size of the training database. Finally, multi-fidelity ROMs which can predict the entire flow field are investigated: graph neural networks [2] are trained to learn the difference between a low-fidelity field (obtained by a simulation on a coarse mesh) and a high-fidelity field (obtained by a simulation on a fine mesh). In this way, efficient and accurate predictions can be obtained by performing simulations on coarse grids and then post-processing the results through the graph neural network which corrects the low-fidelity results. REFERENCES [1] M. Giselle Fernández-Godino. Review of multi-fidelity models. Advances in Computational Science and Engineering, 2023, 1(4): 351-400. [2] Zhou, J., Cui, G., Hu, S., Zhang, Z., Yang, C., Liu, Z., ... & Sun, M. (2020). Graph neural networks: A review of methods and applications. AI open, 1, 57-81.
