Physics latent manifold-guided flow fields reconstruction for accelerating supersonic combustion simulations

Accelerating convergence has long been an important objective in computational fluid dynamics (CFD). Recent advances in machine learning provide new approaches to accelerating numerical simulations. This study proposes a physics latent manifold framework for reconstructing supersonic combustion flow fields and provides high-quality initial conditions for open-source solvers. The proposed supersonic combustion flow reconstruction neural network (SCNN) adopts a two-stage encoder-decoder architecture. Through data-driven encoding, SCNN implicitly embeds multimodal physics information, including geometric parameters and boundary conditions onto a low-dimensional latent manifold, then reconstructs detailed three-dimensional multiphysics fields from the latent manifold. For the tested hydrogen-fueled cavity combustors, SCNN initialization provides an average acceleration of approximately two while retaining good agreement with the converged CFD solutions. The reconstructed initial fields also improve convergence robustness for cases that exhibit poor convergence under conventional initialization. By combining the efficiency of deep learning with the robustness of a conventional numerical solver, the proposed framework provides a practical approach to accelerating repeated supersonic-combustion simulations.