Statistical-learning method for predicting hydrodynamic drag, lift, and pitching torque on spheroidal particles.

Area of Science:

• Fluid dynamics
• Computational science
• Statistical learning

Background:

• Predicting hydrodynamic interactions of non-spherical particles is complex.
• Existing empirical correlations have limitations in high-dimensional input spaces.

Purpose of the Study:

• To develop a statistical learning approach for predicting hydrodynamic interactions of thin oblate spheroidal particles.
• To replace conventional empirical correlations with a more accurate neural-network-based model.

Main Methods:

• Performed resolved simulations of steady uniform flow around a 1:10 spheroidal body (1 ≤ Re ≤ 120).
• Collected a database of Reynolds number- and orientation-dependent drag, lift, and pitching torque.
• Trained and validated a multilayer perceptron using the generated database.

Main Results:

• The neural network accurately predicts hydrodynamic forces (drag, lift, torque).
• The statistical approach shows higher accuracy than existing empirical correlations.
• Simulations of buoyancy-driven disk motion using the model show good agreement with experimental data.

Conclusions:

• The neural network-based statistical learning approach offers a more accurate prediction of hydrodynamic interactions.
• This method has significant potential for improving particle-resolved simulations, especially for non-spherical particles.
• The findings validate the method's effectiveness against experimental observations.