Research Of Supersonic Mixing Layer Flow Based On Machine Learning

Based on effective mixing of fuel and air in RBCC and cross application of machine learning method in flow study,machine learning is adapted to supersonic mixing layer flow for estimation of flow parameters,analysis of turbulent tensor field,extraction of interesting zone and research of dynamic characteristics of fluid system.Research progress on cold flow and mixing enhancement technology of supersonic mixing layer is simply introduced.Then a detailed report on the application of machine learning,especially newly developing deep learning,in flow research is reviewed from four aspects,which are prediction and estimation of flow parameters and flow field structure,flow feature extraction and flow model recognition,computational fluid dynamics,as well as control of nonlinear turbulent flow.Increasing characteristics of supersonic mixing layer is studied,and change of mixing layer thickness on steamwise direction as well as variation of increasing rate with velocity ratio and density ratio is considered.Then some comments and discussion on empirical formula of supersonic mixing layer increasing rate are made and a prediction and estimation model based on deep learning is proposed.The study results indicate that there exist strongly nonlinear relationship between increasing rate and velocity ratio as well as density ratio,and the newly developed deep model,with good generalizing ability,is pretty accurate and may be used for optimization of mixing enhancement device.Turbulent tensor field is studied by means of K-Means cluster method,and variation of cluster centroids with cluster number,spatial distribution of different clusters as well as temporal evolution of cluster centroids are discussed.The study results show that this method can recognize major turbulent tensor model and different clusters,with different spatial distribution characteristics,may represent different flow feature.Besides,the study result of temporal evolution indicate that cluster centroids may appropriately keep constant while the structure of flow field may vary with time for fully developed supersonic mixing layer flow.An anlysis method of dynamic characteristics of flow system based on K-Means cluster and Markov chain model is simply introduced and is adapted to the research of supersonic mixing layer flow.Homogeneousness of clusters,characteristics of condition transition and ergodicity of flow system are considered.The method is explained from the opinion of continuous differential dynamic system and is compared with POD method.The study results indicate that the clusters are homogeneous and variation of clustering condition with time step displays quite randomness.The fluid system of supersonic mixing layer,for which the average of sample can be considered as that of ensemble,is ergodic.And the partial differential fluid system with variant coefficients is transferred to an autonomous original differential system with constant coefficients by K-M model,which may be considered as a flow reduced-order model.Proper orthogonal decomposition method is applied for supersonic mixing layer flow,and distribution of modes energy,temporal evolution and frequency characteristics of mode coefficient,as well as spatial structure of different modes are studied.The result indicate that there exist a relatively concentrated energy distribution for modes of different flow parameters.For streamwise velocity,the first two modes,with isolines symmetry around centerline in first mode and isolines mostly in the side of low velocity in second modes,mainly reflect mean flow characteristics,while other modes fluctuation characteristics.Besides,frequency of streamwise velocity mode coefficient increases with velocity ratio.For transverse velocity,there exist some conjugate modes,whose modes coefficients,with constant phase difference,vary similarly with time and spatial structures are staggered.First four modes,with relatively regular spatial structure,may reflect some large scale characteristics.Besides,eigen frequency of transverse velocity modes increase with the increase of mode number.POD shows relatively good reduced-order effect for supersonic mixing layer,and POD approximations are close to numerical simulation results for each flow parameters.However,there still exists some difference between approximations and simulation results,which is much more apparent in mixing zone around centerline with great parameter gradient.Besides,the approximate error decreases with the increase of the number of approximate modes.