A method to estimate the potential for non-linear compression in turbulent flows
The local tangent residual helps determine if non-linear dimension reduction is worthwhile for specific datasets.
Researchers proposed the local tangent residual (LTR) as an a priori estimate. This estimate measures the potential benefit of non-linear dimension reduction for a given dataset.
The LTR is based on the local geometry of the data points. It requires neither prior model training nor the definition of a threshold parameter.
The study investigated two vortex street cases and the turbulent wake flow of a wind turbine. In these cases, the LTR reproduced dominant trends in autoencoder compression performance.
The decoder linearisation error (DLE) quantifies the non-linearity of a trained decoder. It compares the decoder to a linearised surrogate model to enable physical interpretability.
A large non-linear compression potential in turbulent flow is corroborated by a large DLE. This means a pronounced non-linearity was captured in the trained autoencoder.
The LTR helps practitioners determine if non-linear compression is worthwhile. This avoids the substantial computational cost of autoencoder architecture search and hyperparameter tuning when proper orthogonal decomposition (POD) already suffices.
Key facts
- The local tangent residual (LTR) uses local geometry of data points to estimate compression potential.
- The LTR does not require prior model training or a threshold parameter.
- The decoder linearisation error (DLE) quantifies non-linearity by comparing a decoder to a linearised surrogate model.
- The method was tested on two vortex street cases and a wind turbine turbulent wake flow.