(701b) Machine Learning Based Interaction Force Model for Non-Spherical Particles in Incompressible Flows | AIChE

(701b) Machine Learning Based Interaction Force Model for Non-Spherical Particles in Incompressible Flows


Hwang, S. - Presenter, The Ohio State University
Pan, J., Ohio State University
Fan, L. S., Ohio State University
Gas solid flow systems are ubiquitous in a variety of applications such as pneumatic transport, combustion of pulverized coal, drying, sand blasting, and combustion of chemical looping systems. Computational fluid dynamics (CFD) has been an effective way to predict the particulate flow and to design the system at a large-scale with the methodologies such as the discrete element method (DEM) and two fluid method. In applications, the interaction force model between gas and non-spherical particles is required, but it is hard to formulate due to the complicated fluid forces which are shape- and orientation-dependent. The objective of this study is to develop the force model for non-spherical particles using the neural network approach which is advantageous to correlate the complicated relationship between variables.

This study uses particle resolved direct numerical simulation (PR-DNS) to collect the datasets for the interaction force such as drag force, lifting force and torque between a particle and an incompressible flow from a low to moderate Reynolds number. The PR-DNS is based on the simplified gas kinetic scheme coupled with the immersed boundary method and the non-spherical particles having different roughness, aspect ratio and orientation are generated by using spherical harmonic function. This approach has the advantage of being more accurate compared to the experimental method because it can reflect the exact shape of the particle and calculate the forces. To develop the force model, we utilize a variational auto-encoder to extract the geometrical features affecting the gas-solid interaction. Furthermore, we apply an artificial neural network to correlate the geometrical features and the flow conditions with the interaction force from the PR-DNS method. The preliminary results using 2,800 datasets show the mean absolute percentage error of 15% for the drag force coefficients, which is lower than 27% when the spherical particles are assumed. This study will provide a drag force model which can be coupled with DEM and consider the high nonlinearity of shape factors, orientation, viscosity, and flow intensity. It will also give a correlation for the lifting force that is important for some cases involving asymmetrical flow around the particles having high aspect ratio. By implementing the force model for non-spherical particles in CFD, this study will improve the accuracy of the multi-phase simulation and utility for a wide range of industrial designs and applications.