Overview
This research explored compressible flow past a rapidly rotating cylinder through high-fidelity numerical simulations. The study focused on understanding the evolution of aerodynamic loads and flow instability within a Reynolds number (Re) range of 1000 to 6000. It also investigated the application of supervised machine learning techniques, such as polynomial regression, Bayesian regression, and artificial neural networks (ANNs), to model the observed highly nonlinear dependencies in the flow data.
Research Context
The investigation of compressible flow around a rotating cylinder involves complex fluid dynamics. High-fidelity numerical simulations are resource-intensive, requiring 1 million core hours for the 101 simulations in this study. The goal was to understand the intricate transitions in flow behavior, particularly from periodic vortex shedding to complex multi-mode oscillatory states, and to develop data-driven models capable of capturing these nonlinear characteristics efficiently.
Approach
The researchers utilized high-fidelity numerical simulations to generate a dataset comprising 101 data points. This dataset provided detailed information on aerodynamic loads and flow instability across the specified Reynolds number range. Spectral analysis was applied to lift and drag signals to identify dominant frequencies and analyze their interactions, including amplitude modulation and nonlinear mode coupling. Three primary machine learning approaches were sequentially investigated:
- Polynomial Regression: This method served as a baseline to fit the observed data.
- Bayesian Regression: Frameworks employing B-spline and Gaussian radial basis functions were used to enhance flexibility and provide uncertainty quantification. Spline-based models were specifically evaluated for their ability to capture piecewise nonlinear trends.
- Artificial Neural Networks (ANNs): ANNs were developed as high-capacity surrogate models. Their performance was assessed for predicting maximum lift coefficient, instability onset time, and drag coefficient. The ANN was further evaluated as a generative model for reconstructing flow behavior at unseen Reynolds numbers. A hierarchical refinement strategy was incorporated into the ANN-based models.
Findings
- The numerical simulations revealed a transition in flow behavior from periodic vortex shedding to complex multi-mode oscillatory states.
- A critical bifurcation was identified near Re = 5650.
- Spectral analysis of lift and drag signals indicated the emergence and interaction of multiple dominant frequencies.
- Post-bifurcation, amplitude modulation and nonlinear mode coupling were observed in the flow.
- Polynomial regression provided baseline fits but demonstrated limitations in capturing localized fluctuations, especially near the bifurcation point.
- Bayesian regression frameworks, particularly those using B-splines, showed improved flexibility and uncertainty quantification, demonstrating superior performance in capturing piecewise nonlinear trends.
- Artificial neural networks achieved excellent predictive accuracy for the maximum lift coefficient and instability onset time.
- While ANNs maintained reasonable fidelity for the more challenging drag coefficient, their performance for this specific metric was not characterized as 'excellent'.
- When trained on high-fidelity data, ANN-based models were shown to function as efficient and reliable surrogates for complex fluid dynamics problems, and could be used as generative models to reconstruct flow behavior at unseen Re with a hierarchical refinement strategy.
Why This Matters
The development of efficient and reliable surrogate models, particularly ANNs trained on high-fidelity data, offers a pathway to analyzing complex fluid dynamics problems with reduced computational expense. Understanding and predicting nonlinear flow behaviors, such as those observed in compressible flow past rotating cylinders, can contribute to areas where such phenomena are critical.