Inferring turbulent environments via machine learning.

Area of Science:

• Turbulence research
• Fluid dynamics
• Astrophysics
• Geophysics
• Machine learning applications

Background:

• Classifying turbulent environments from partial observations is vital for engineering, earth observation, and astrophysics.
• Existing methods often require system knowledge or extensive data, which may not always be available.
• This study addresses the challenge in a model-free setting with ample, high-quality data.

Purpose of the Study:

• To classify turbulent environments using partial observational data in a model-free scenario.
• To compare the effectiveness of a deep convolutional neural network (DCNN) against Bayesian inference.
• To identify key physical features utilized by the DCNN for classification through an ablation study.

Main Methods:

• Utilized 10 turbulent 'ensembles' generated by varying rotation frequency in a 3D domain.
• Employed partial observations limited to instantaneous kinetic energy distribution in a 2D plane.
• Compared a state-of-the-art deep convolutional neural network (DCNN) with Bayesian inference methods.

Main Results:

• The machine learning (ML) approach using DCNN demonstrated superior performance compared to Bayesian inference.
• ML performance remained robust across variations in training data quantity and hyper-parameter tuning.
• An ablation study successfully ranked the importance of flow features, revealing key physical insights used by the DCNN.

Conclusions:

• Deep convolutional neural networks offer a powerful, data-driven solution for classifying turbulent flows from partial observations.
• The DCNN approach is effective even with limited training data, showcasing its adaptability.
• Further research into data-driven methods can unlock significant applications in various scientific and engineering fields.