Predicting the Steady Flow of a Fluid with Particles by Deep Learning

Computational Fluid Dynamics (CFD) simulation has the potential for application in material science. In these applications, a common study object is a highly viscous fluid that passes structures in the microscale. However, its computational cost is a barrier to application. Meanwhile, some recent studies successfully reduced the computational cost of CFD by utilizing machine learning and deep learning techniques [1, 2].
Using deep learning, the authors predicted the steady flow around a large number of particles and examined the effectiveness of the prediction for the study of materials development. The particle-fluid interaction is computed using the Smoothed Profile Method [3]. After learning the flow that passes the particles obtained by iterative calculation, the deep learning quickly and accurately predicted the flow of the system with unknown particle concentration and arrangement. The fluid force applied to each particle was also accurately predicted.

References
[1] X. Guo, W. Li, and F. Iorio, KDD ’16 (2016).
[2] O. Hennigh, arXiv, arXiv:1710.10352 (2017).
[3] Y. Nakayama and R. Yamamoto, Phys. Rev. E, 71, 036707 (2005).