Dynamic mode decomposition (DMD) and deep learning are data-driven approaches that allow a description of the target phenomena in new representation spaces. This fact motivates their comparison in the analysis of flow data, generated through experimental setups and numerical simulations. The focused application is the processing of high-speed videos of horizontal two-phase stratified and slug flows regimes. Henceforth, in this work, we consider the traditional DMD, the sparsity-promoting DMD (SPDMD) and, in the deep learning context, we select an unsupervised convolutional autoencoder (CAE). In this avenue, it becomes imperative to compare DMD and deep learning with respect to: computational complexity of target techniques; reduced order modeling versus data representation; data set necessary to compute the dynamic modes and deep learning training; the preservation of the phase interface in the DMD and CAE space; data synthesis. In general, the results favor DMD in the considered applications.

中文翻译：

---

两相流分析的动态模态分解与深度学习技术的比较

动态模式分解 (DMD) 和深度学习是数据驱动的方法，可以在新的表示空间中描述目标现象。这一事实促使他们在流量数据分析中进行比较，这些数据是通过实验设置和数值模拟生成的。重点应用是处理水平两相分层和段塞流状态的高速视频。此后，在这项工作中，我们考虑了传统的 DMD、稀疏促进 DMD（SPDMD），并且在深度学习环境中，我们选择了一个无监督的卷积自动编码器（CAE）。在这条道路上，必须在以下方面比较 DMD 和深度学习：目标技术的计算复杂性；降阶建模与数据表示；计算动态模式和深度学习训练所需的数据集；在 DMD 和 CAE 空间中保留相界面；数据合成。一般来说，结果有利于 DMD 在所考虑的应用中。