We establish a deconvolutional artificial-neural-network(D-ANN)approach in large-eddy simulation(LES)of compressible turbulent flow.Filtered variables in the neighboring locations are taken as the inputs of D-ANN to r...We establish a deconvolutional artificial-neural-network(D-ANN)approach in large-eddy simulation(LES)of compressible turbulent flow.Filtered variables in the neighboring locations are taken as the inputs of D-ANN to recover original(unfiltered)variables,including density,momentum and pressure.The scale-similarity form is adopted to reconstruct subfilter-scale(SFS)terms.The proposed D-ANN models can give better a priori predictions of the sub-filter stress and heat flux than the classical approximate-deconvolution method(ADM)and the velocity-gradient model(VGM).The predicted SFS terms with the D-ANN models have correlation coefficients larger than 98.4%and relative errors smaller than 18%.In the a posteriori analysis,the D-ANN model compares against the implicit LES(ILES),the dynamic-Smagorinsky model(DSM),and the dynamic-mixed model(DMM).The D-ANN model predicts better than these classical models for velocity spectra,statistical properties of SFS kinetic energy flux and velocity increments.The turbulence statistics and transient velocity divergence are also accurately reconstructed.The type of explicit filter and the impact of compressibility do not significantly affect a posteriori accuracy of the D-ANN model.Results showthat the proposed D-ANN approach has a great potential in developing highly accurate SFS models for large-eddy simulation of complex compressible turbulent flow.展开更多
基金This research was supported by the National Nat542 ural Science Foundation of China(Grants 91952104,92052301 and 91752201).
文摘We establish a deconvolutional artificial-neural-network(D-ANN)approach in large-eddy simulation(LES)of compressible turbulent flow.Filtered variables in the neighboring locations are taken as the inputs of D-ANN to recover original(unfiltered)variables,including density,momentum and pressure.The scale-similarity form is adopted to reconstruct subfilter-scale(SFS)terms.The proposed D-ANN models can give better a priori predictions of the sub-filter stress and heat flux than the classical approximate-deconvolution method(ADM)and the velocity-gradient model(VGM).The predicted SFS terms with the D-ANN models have correlation coefficients larger than 98.4%and relative errors smaller than 18%.In the a posteriori analysis,the D-ANN model compares against the implicit LES(ILES),the dynamic-Smagorinsky model(DSM),and the dynamic-mixed model(DMM).The D-ANN model predicts better than these classical models for velocity spectra,statistical properties of SFS kinetic energy flux and velocity increments.The turbulence statistics and transient velocity divergence are also accurately reconstructed.The type of explicit filter and the impact of compressibility do not significantly affect a posteriori accuracy of the D-ANN model.Results showthat the proposed D-ANN approach has a great potential in developing highly accurate SFS models for large-eddy simulation of complex compressible turbulent flow.
