Dehong Fang, Jifu Tan, Immersed boundary-physics informed machine
learning approach for fluid–solid coupling, Ocean Engineering, Volume
263, 2022, 112360, ISSN 0029-8018,
https://doi.org/10.1016/j.oceaneng.2022.112360.Abstract: Fluid–solid coupling is commonly used but sometimes
expensive in large-scale simulations for fluid dynamics. Conventional
numerical methods rely on high performance computers and parallel
computing techniques to accelerate simulations. In this work, a
lightweight immersed boundary-physics informed machine learning model
is proposed for the fluid–solid coupling based on the physical
framework of multi-direct forcing of the immersed boundary method. Two
dimensional flows past a static cylinder are adopted as case studies
for the drag. It shows close agreements of drag coefficient among
simulations conducted by the immersed boundary-lattice Boltzmann
method (IB-LBM), immersed boundary-physics informed neural network
model (IB-PINN), and data from references. No-slip boundary conditions
around the cylinder boundaries are closely satisfied and the time
consumed by the machine learning model is reduced by 38.5% compared
with IB-LBM, which demonstrates that the machine learning approach is
robust, fast, and accurate. Keywords: Fluid–solid interaction; Machine
learning; Immersed boundary method; Lattice Boltzmann method
在之前的工作里，我尝试用机器学习模型来取代浸入边界法，先挂到了preprint上。后来发现其他学者用机器学习模型来实现fluid-particle interaction，发在了JFM上。其实我觉得这些问题的本质都差不多，基本上都是利用机器学习模型取代一些复杂的计算过程。
不过在这篇论文里，不管怎么调mesh size，得到的Cd总是比文献中的大，估计选取方形流场来计算Cd不是很精确。
如果不是这些琐碎的问题和我想论文早点发表，我还是想投稿JCP试试，但最后还是选择了Ocean Engineering。
当下CFD有很多和PINN结合的应用，hidden fluid mechanics可以通过给定坐标，时间和定义ns方程的loss来预测流场，其他机器学习模型还可以用来预测NACA foil的各种系数。但是本质上，这些工作都是用机器学习模型来替代一些复杂算法。那么理论上，任何与流场模拟后处理的分析工作都可以用机器学习来替代。例如计算流场管壁的wall shear stress，获得流场中的vortex structures等。因为它们都是基于流场中的数据，额外计算得到结果，都是可以通过机器学习模型来实现。希望这些insights能够帮助到其他同学。