DeepONet-assisted DeepMMnet for full-field reconstruction of Rayleigh–Taylor and tilted-rig turbulent mixing

This study addresses the problem of multi-component turbulent mixing, where traditional numerical methods often suffer from limited generalization capability and inefficient utilization of measurement data. To overcome these limitations, we developed intelligent models tailored for different scenarios. Based on high-fidelity numerical simulation data of Rayleigh–Taylor (RT) mixing, neural operator networks were constructed for the one-dimensional RT turbulent mixing problem and the two-dimensional tilted-rig problem, achieving efficient predictions with relative errors within 5%. For more challenging cases with only sparse mass fraction data, we proposed a deep multi-physical-field neural network (DeepMMnet), which enables accurate reconstruction of the full flow field, including density, velocity, and energy, from limited mass fraction data. Furthermore, predictions under noisy data were investigated to evaluate the robustness of the proposed model under more realistic measurement conditions. For noisy inputs, DeepMMnet successfully reconstructed the full flow field with relative L2 errors below 5%. Overall, the intelligent models developed in this study overcome key limitations of traditional numerical simulation methods and provide new tools and perspectives for accurate simulation and reconstruction of engineering mixing problems.