Application of generative deep learning to predict temperature, flow and species distributions using simulation data of a methane combustor

摘要

A data-driven surrogate modelling methodology of CFD simulation data is proposed.It uses convolutional variational autoencoders (VAE) and multi-layer perceptron (MLP) neural networks in an integrated manner to predict temperature, velocity and species mass fraction profiles on a cell-by-cell basis. The VAE performs feature extraction and data compression, while significantly reducing the number of network interconnections and ensuring physically realistic predictions.The MLP maps the CFD boundary condition values to the encodings generated with the VAE encoder network. The approach is demonstrated via application to a 2D axisymmetric methane-fired turbulent jet diffusion flame. The integrated network model produced average normalized mean absolute errors (NMAE) of 2.41% for the temperature predictions, 0.99% for velocity and 0.925% for species mass fractions. It is, therefore, possible to predict 2D CFD data fields with reasonable accuracy and generalizability, although high NMAE values were observed for certain cells confined to a small region near the centreline of the burner. The methodology lays the foundation for application to larger industrial problems to visualize processes inside equipment, deploy virtual sensors, perform quick what-if analysis, explore the design space, link to optimization routines to effectively control equipment, detect anomalies, or to form part of lower-dimensional system simulations.