The nonlinear evolution of multimode Rayleigh–Taylor (RT) mixing plays a crucial role in numerous natural phenomena and engineering applications, yet accurate characterization of its small-scale dynamics is severely restricted by limited spatial resolution in experiments and simulations. This resolution constraint motivates the development of superresolution (SR) techniques to reconstruct high-frequency flow features from coarsely resolved data. In this study, direct numerical simulations reveal that while horizontally averaged profiles and integral mixing width remain largely insensitive to resolution degradation, fluctuation variance ${\sigma }^2$ decreases sharply once grid spacing exceeds the interfacial thickness, resulting in systematic overestimation of mixedness $\theta$. To recover the lost multiscale information, we develop three convolutional neural network (CNN)-based superresolution models within a unified framework: a baseline CNN architecture, a residual-block variant (Res), and a physically conditional residual variant $(\text{Re}\mathrm{s}+\mathrm{h})$. These models substantially outperform bicubic interpolation. For spatial $\times 4$ and $\times 8$ upscaling of flows with varying initial perturbation phases, the CNN-based approaches reduce relative $L_1$ errors from 6.3% to 0.41% $(\times 4)$ and from 15.1% to approximately 4.5% $(\times 8)$. Furthermore, models trained solely on initial perturbation phase variations exhibit strong generalization to unseen perturbation amplitudes, interface thicknesses, and Reynolds numbers, maintaining high fidelity for the $\times 4$ SR task across parameter space. For the more demanding $\times 8$ task, minor oversharpening and spurious small-scale structures appear under certain conditions. These artifacts are effectively suppressed by enriching the training set with additional cases spanning amplitudes, thicknesses, and Reynolds numbers. These results highlight the strong potential of physics-informed machine learning approaches for exploring more complex three-dimensional RT mixing problems.

Close

PrevNext

View figure in article

• Hydrodynamic instabilities
• Rayleigh-Taylor instability
• Machine learning